Depth estimation in videos is essential for visual perception in real-world applications. However, existing methods either rely on simple frame-by-frame monocular models, leading to temporal inconsistencies and inaccuracies, or use computationally demanding temporal modeling, unsuitable for real-time applications. These limitations significantly restrict general applicability and performance in practical settings. To address this, we propose VeloDepth, an efficient and robust online video depth estimation pipeline that effectively leverages spatiotemporal priors from previous depth predictions and performs deep feature propagation. Our method introduces a novel Propagation Module that refines and propagates depth features and predictions using flow-based warping coupled with learned residual corrections. In addition, our design structurally enforces temporal consistency, resulting in stable depth predictions across consecutive frames with improved efficiency. Comprehensive zero-shot evaluation on multiple benchmarks demonstrates the state-of-the-art temporal consistency and competitive accuracy of VeloDepth, alongside its significantly faster inference compared to existing video-based depth estimators. VeloDepth thus provides a practical, efficient, and accurate solution for real-time depth estimation suitable for diverse perception tasks. Code and models are available at github.com/lpiccinelli-eth/velodepth.

Recent advances in Multimodal Large Language Models (MLLMs) have significantly pushed the frontier of egocentric video question answering (EgocentricQA). However, existing benchmarks and studies are mainly limited to common daily activities such as cooking and cleaning. In contrast, real-world deployment inevitably encounters domain shifts, where target domains differ substantially in both visual style and semantic content. To bridge this gap, we introduce EgoCross, a comprehensive benchmark designed to evaluate the cross-domain generalization of MLLMs in EgocentricQA. EgoCross covers four diverse and challenging domains, including surgery, industry, extreme sports, and animal perspective, representing realistic and high-impact application scenarios. It comprises approximately 1,000 QA pairs across 798 video clips, spanning four key QA tasks: prediction, recognition, localization, and counting. Each QA pair provides both OpenQA and CloseQA formats to support fine-grained evaluation. Extensive experiments show that most existing MLLMs, whether general-purpose or egocentric-specialized, struggle to generalize to domains beyond daily life, highlighting the limitations of current models. Furthermore, we conduct several pilot studies, e.g., fine-tuning and reinforcement learning, to explore potential improvements. We hope EgoCross and our accompanying analysis will serve as a foundation for advancing domain-adaptive, robust egocentric video understanding.

Differentiable simulators represent an environment’s dynamics as a differentiable function. Within robotics and autonomous driving, this property is used in Analytic Policy Gradients (APG), which relies on backpropagating through the dynamics to train accurate policies for diverse tasks. Here we show that differentiable simulation also has an important role in world modeling, where it can impart predictive, prescriptive, and counterfactual capabilities to an agent. Specifically, we design three novel task setups in which the differentiable dynamics are combined within an end-to-end computation graph not with a policy, but a state predictor. This allows us to learn relative odometry, optimal planners, and optimal inverse states. We collectively call these predictors Analytic World Models (AWMs) and demonstrate how differentiable simulation enables their efficient, end-to-end learning. In autonomous driving scenarios, they have broad applicability and can augment an agent’s decision-making beyond reactive control.

Planning allows an agent to safely refine its actions before executing them in the real world. In autonomous driving, this is crucial to avoid collisions and navigate in complex, dense traffic scenarios. One way to plan is to search for the best action sequence. However, this is challenging when all necessary components – policy, next-state predictor, and critic – have to be learned. Here we propose Differentiable Simulation for Search (DSS), a framework that leverages the differentiable simulator Waymax as both a next state predictor and a critic. It relies on the simulator’s hardcoded dynamics, making state predictions highly accurate, while utilizing the simulator’s differentiability to effectively search across action sequences. Our DSS agent optimizes its actions using gradient descent over imagined future trajectories. We show experimentally that DSS – the combination of planning gradients and stochastic search – significantly improves tracking and path planning accuracy compared to sequence prediction, imitation learning, model-free RL, and other planning methods.

State-of-the-art semantic segmentation models are typically optimized in a data-driven fashion, minimizing solely per-pixel or per-segment classification objectives on their training data. This purely data-driven paradigm often leads to absurd segmentations, especially when the domain of input images is shifted from the one encountered during training. For instance, state-of-the-art models may assign the label "road" to a segment that is included by another segment that is respectively labeled as "sky". However, the ground truth of the existing dataset at hand dictates that such inclusion is not feasible. Our method, Infeasible Semantic Inclusions (InSeIn), first extracts explicit inclusion constraints that govern spatial class relations from the semantic segmentation training set at hand in an offline, data-driven fashion, and then enforces a morphological yet differentiable loss that penalizes violations of these constraints during training to promote prediction feasibility. InSeIn is a light-weight plug-and-play method, constitutes a novel step towards minimizing infeasible semantic inclusions in the predictions of learned segmentation models, and yields consistent and significant performance improvements over diverse state-of-the-art networks across the ADE20K, Cityscapes, and ACDC datasets. Codebase will be made available. Code is available at https://github.com/SHAMIK-97/InSeIn

Visual prompt-based detection enables generalization to arbitrary novel instances in the target image by using one or a few visual templates. Previous methods rely on complex relation or explicit feature matching modules, and their designs are deeply coupled with specific detectors, greatly limiting their applicability. Instead, we propose our ‘Any Detector can Detect Anything’ framework that can enable any detector to detect any object given a single or a few visual templates. Specifically, we design an adapter called Template-Aware Adapter that can be added on top of any existing detector architecture to inject visual template information directly into the detection features. After integration, localization is done on the feature maps as in standard object detectors, effectively transforming any detector into a visual prompt-based detector. Furthermore, we revisit current visual prompt detection benchmarks and correct their unrealistic test assumptions and class splits, which limit the usability of the developed algorithms in the real world. We introduce a set of realistic benchmarks to remedy these issues. We comprehensively evaluate the proposed model on both existing and our new benchmarks, outperforming current state-of-the-art one-shot and few-shot detection methods by a large margin.

Federated learning (FL) has advanced semantic segmentation through decentralized training to reduce annotation costs. However, most FL-based semantic segmentation methods assume fixed foreground classes, resulting in catastrophic forgetting of old categories when local clients continually collect streaming data of new classes without storing old categories. Moreover, the irregular participation of new local clients with novel classes unseen by others may exacerbate heterogeneous forgetting across clients during global FL training. To resolve the above challenges, we propose a Hierarchical Forgetting Alleviation (HFA) model. By tackling forgetting within and across local clients, our model ensures that all local clients learn from each other as they continuously learn new categories. Specifically, to alleviate class-imbalanced forgetting within local clients induced by background shift, we develop a confidence-regularized pseudo labeling strategy to produce class-balanced soft pseudo labels for old categories that are labeled as background. Guided by soft pseudo labels, we design a graph-induced relation matching loss and a forgetting-balanced gradient propagation module to tackle ambiguous inter-class relations and class-imbalanced gradient propagation among old classes. Besides, a novel task detection module and an adaptive DBSCAN clustering are devised to address inter-client heterogeneous forgetting. They detect the arrival of new tasks to store the old global model for local pseudo labeling and distillation, while supplying global class prototypes for modeling inter-class relations and warm-starting global classifier. Experiments on multiple datasets verify our model's superiority over other methods.

Accurate monocular metric depth estimation (MMDE) is crucial to solving downstream tasks in 3D perception and modeling. However, the remarkable accuracy of recent MMDE methods is confined to their training domains. These methods fail to generalize to unseen domains even in the presence of moderate domain gaps, which hinders their practical applicability. We propose a new model, UniDepthV2, capable of reconstructing metric 3D scenes from solely single images across domains. Departing from the existing MMDE paradigm, UniDepthV2 directly predicts metric 3D points from the input image at inference time without any additional information, striving for a universal and flexible MMDE solution. In particular, UniDepthV2 implements a self-promptable camera module predicting a dense camera representation to condition depth features. Our model exploits a pseudo-spherical output representation, which disentangles the camera and depth representations. In addition, we propose a geometric invariance loss that promotes the invariance of camera-prompted depth features. UniDepthV2 improves its predecessor UniDepth model via a new edge-guided loss which enhances the localization and sharpness of edges in the metric depth outputs, a revisited, simplified and more efficient architectural design, and an additional uncertainty-level output which enables downstream tasks requiring confidence. Thorough evaluations on ten depth datasets in a zero-shot regime consistently demonstrate the superior performance and generalization of UniDepthV2. Code and models are available at: github.com/lpiccinelli-eth/UniDepth.

Level-5 driving automation requires a robust visual perception system that can parse input images under any condition. However, existing driving datasets for dense semantic perception are either dominated by images captured under normal conditions or are small in scale. To address this, we introduce ACDC, the Adverse Conditions Dataset with Correspondences for training and testing methods for diverse semantic perception tasks on adverse visual conditions. ACDC consists of a large set of 8012 images, half of which (4006) are equally distributed between four common adverse conditions: fog, nighttime, rain, and snow. Each adverse-condition image comes with a high-quality pixel-level panoptic annotation, a corresponding image of the same scene under normal conditions, and a binary mask that distinguishes between intra-image regions of clear and uncertain semantic content. 1503 of the corresponding normal-condition images feature panoptic annotations, raising the total annotated images to 5509. ACDC supports the standard tasks of semantic segmentation, object detection, instance segmentation, and panoptic segmentation, as well as the newly introduced uncertainty-aware semantic segmentation. A detailed empirical study demonstrates the challenges that the adverse domains of ACDC pose to state-of-the-art supervised and unsupervised approaches and indicates the value of our dataset in steering future progress in the field.

3D Gaussian Splatting (3DGS) has attracted great attention in novel view synthesis because of its superior rendering efficiency and high fidelity. However, the trained Gaussians suffer from severe zooming degradation due to non-adjustable representation derived from single-scale training. Though some methods attempt to tackle this problem via post-processing techniques such as selective rendering or filtering techniques towards primitives, the scale-specific information is not involved in Gaussians. In this paper, we propose a unified optimization method to make Gaussians adaptive for arbitrary scales by self-adjusting the primitive properties (e.g., color, shape and size) and distribution (e.g., position). Inspired by the mipmap technique, we design pseudo ground-truth for the target scale and propose a scale-consistency guidance loss to inject scale information into 3D Gaussians. Our method is a plug-in module, applicable for any 3DGS models to solve the zoomin and zoom-out aliasing. Extensive experiments demonstrate the effectiveness of our method. Notably, our method outperforms 3DGS in PSNR by an average of 9.25 dB for zoom-in and 10.40 dB for zoom-out on NeRF Synthetic dataset. Our project website: https://github.com/renaissanceee/Mipmap-GS.

3D Gaussian Splatting (3DGS) has become the de facto method of 3D representation in many vision tasks. This calls for the 3D understanding directly in this representation space. To facilitate the research in this direction, we first build a large-scale dataset of 3DGS using the commonly used ShapeNet and ModelNet datasets. Our dataset ShapeSplat consists of 65K objects from 87 unique categories, whose labels are in accordance with the respective datasets. The creation of this dataset utilized the computing equivalent of 2 GPU years on a TITAN XP GPU. We utilize our dataset for unsupervised pretraining and supervised finetuning for classification and segmentation tasks. To this end, we introduce Gaussian-MAE, which highlights the unique benefits of representation learning from Gaussian parameters. Through exhaustive experiments, we provide several valuable insights. In particular, we show that (1) the distribution of the optimized GS centroids significantly differs from the uniformly sampled point cloud (used for initialization) counterpart; (2) this change in distribution results in degradation in classification but improvement in segmentation tasks when using only the centroids; (3) to leverage additional Gaussian parameters, we propose Gaussian feature grouping in a normalized feature space, along with splats pooling layer, offering a tailored solution to effectively group and embed similar Gaussians, which leads to notable improvement in finetuning tasks. Our dataset and model are publicly available at ShapeSplat.

Human activities are inherently complex, and even simple household tasks involve numerous object interactions. To better understand these activities, it is crucial to model their interactions with the environment captured through dynamic changes. The recent availability of affordable head-mounted cameras and egocentric data offers a more accessible and efficient means to understand dynamic human-object interactions in 3D environments. However, most existing methods for human activity modeling either focus on reconstructing 3D models of hand-object or human-scene interactions or on mapping 3D scenes, neglecting dynamic interactions with objects. The few existing solutions often require inputs from multiple sources, including multi-camera setups, depth-sensing cameras, or kinesthetic sensors. To this end, we introduce EgoGaussian, the first method capable of simultaneously reconstructing 3D scenes and dynamically tracking 3D object motion from RGB egocentric input alone. We leverage the uniquely discrete nature of Gaussian Splatting and segment dynamic interactions from the background. Our approach employs a clip-level online learning pipeline that leverages the dynamic nature of human activities, allowing us to reconstruct the temporal evolution of the scene in chronological order and track rigid object motion. Additionally, our method automatically segments object and background Gaussians, providing explicit 3D representations for both static scenes and dynamic objects. EgoGaussian shows significant improvements in terms of both dynamic object and background reconstruction quality compared to the state-of-the-art. We also qualitatively demonstrate the high quality of the reconstructed models.

Foundational vision-language models like CLIP are emerging as a promising paradigm in vision due to their excellent generalization. However, adapting these models for downstream tasks while maintaining their generalization remains challenging. In literature, one branch of methods adapts CLIP by learning prompts using images. While effective, these methods often rely on image-label data, which is not always practical, and struggle to generalize to new datasets due to overfitting on few-shot source data. Another approach explores training-free methods by generating class captions from large language models (LLMs) and performing prompt ensembling, but these methods often produce static, class-specific prompts that cannot be transferred to new classes and incur additional costs by generating LLM descriptions for each class separately. In this work, we aim to combine the strengths of both approaches by learning prompts using only text data derived from LLMs. As supervised training of prompts in the image-free setup is non-trivial, we develop a language-only efficient training approach that enables prompts to distill rich contextual knowledge from LLM data. Furthermore, by mapping the LLM contextual text data within the learned prompts, our approach enables zero-shot transfer of prompts to new classes and datasets, potentially reducing the LLM prompt engineering cost. To the best of our knowledge, this is the first work that learns generalized and transferable prompts for image tasks using only text data. We perform evaluations on 4 benchmarks, where ProText improves over ensembling methods while being competitive with those using labeled images.

Object detection, particularly open-vocabulary object detection, plays a crucial role in Earth sciences, such as environmental monitoring, natural disaster assessment, and land-use planning. However, existing open-vocabulary detectors, primarily trained on natural-world images, struggle to generalize to remote sensing images due to a significant data domain gap. Thus, this paper aims to advance the development of open-vocabulary object detection in remote sensing community. To achieve this, we first reformulate the task as Locate Anything on Earth (LAE) with the goal of detecting any novel concepts on Earth. We then developed the LAE-Label Engine which collects, auto-annotates, and unifies up to 10 remote sensing datasets creating the LAE-1M — the first large-scale remote sensing object detection dataset with broad category coverage. Using the LAE-1M, we further propose and train the novel LAE-DINO Model, the first open-vocabulary foundation object detector for the LAE task, featuring Dynamic Vocabulary Construction (DVC) and Visual-Guided Text Prompt Learning (VisGT) modules. DVC dynamically constructs vocabulary for each training batch, while VisGT maps visual features to semantic space, enhancing text features. We comprehensively conduct experiments on established remote sensing benchmark DIOR, DOTAv2.0, as well as our newly introduced 80-class LAE-80C benchmark. Results demonstrate the advantages of the LAE-1M dataset and the effectiveness of the LAE-DINO method.

6D object pose estimation remains challenging for many applications due to dependencies on complete 3D models, multi-view images, or training limited to specific object categories. These requirements make generalization to novel objects difficult for which neither 3D models nor multi-view images may be available. To address this, we propose a novel method One2Any that estimates the relative 6-degrees of freedom (DOF) object pose using only a single reference-single query RGB-D image, without prior knowledge of its 3D model, multi-view data, or category constraints. We treat object pose estimation as an encoding-decoding process: first, we obtain a comprehensive Reference Object Pose Embedding (ROPE) that encodes an object’s shape, orientation, and texture from a single reference view. Using this embedding, a U-Net-based pose decoding module produces Reference Object Coordinate (ROC) for new views, enabling fast and accurate pose estimation. This simple encoding-decoding framework allows our model to be trained on any pair-wise pose data, enabling large-scale training and demonstrating great scalability. Experiments on multiple benchmark datasets demonstrate that our model generalizes well to novel objects, achieving state-of-the-art accuracy and robustness even rivaling methods that require multi-view or CAD inputs, at a fraction of compute. Code is available at https://github.com/lmy1001/One2Any.

The automated analysis of chemical literature holds promise to accelerate discovery in fields such as material science and drug development. In particular, search capabilities for chemical structures and Markush structures (chemical structure templates) within patent documents are valuable, e.g., for prior-art search. Advancements have been made in the automatic extraction of chemical structures from text and images, yet the Markush structures remain largely unexplored due to their complex multi-modal nature. In this work, we present MarkushGrapher, a multimodal approach for recognizing Markush structures in documents. Our method jointly encodes text, image, and layout information through a Vision-Text-Layout encoder and an Optical Chemical Structure Recognition vision encoder. These representations are merged and used to autoregressively generate a sequential graph representation of the Markush structure along with a table defining its variable groups. To overcome the lack of real-world training data, we propose a synthetic data generation pipeline that produces a wide range of realistic Markush structures. Additionally, we present M2S, the first annotated benchmark of real-world Markush structures, to advance research on this challenging task. Extensive experiments demonstrate that our approach outperforms state-of-the-art chemistry-specific and general-purpose vision-language models in most evaluation settings. Code, models, and datasets are available1.

Monocular 3D estimation is crucial for visual perception. However, current methods fall short by relying on oversimplified assumptions, such as pinhole camera models or rectified images. These limitations severely restrict their general applicability, causing poor performance in real-world scenarios with fisheye or panoramic images and resulting in substantial context loss. To address this, we present UniK3D1, the first generalizable method for monocular 3D estimation able to model any camera. Our method introduces a spherical 3D representation which allows for better disentanglement of camera and scene geometry and enables accurate metric 3D reconstruction for unconstrained camera models. Our camera component features a novel, model-independent representation of the pencil of rays, achieved through a learned superposition of spherical harmonics. We also introduce an angular loss, which, together with the camera module design, prevents the contraction of the 3D outputs for wide-view cameras. A comprehensive zero-shot evaluation on 13 diverse datasets demonstrates the state-of-the-art performance of UniK3D across 3D, depth, and camera metrics, with substantial gains in challenging large-field-of-view and panoramic settings, while maintaining top accuracy in conventional pinhole small-field-of-view domains. Code and models are available at github.com/lpiccinelli-eth/unik3d.

Modern world models require costly and time-consuming collection of large video datasets with action demonstrations by people or by environment-specific agents. To simplify training, we focus on using many virtual environments for inexpensive, automatically collected interaction data. Genie [5], a recent multi-environment world model, demonstrates simulation abilities of many environments with shared behavior. Unfortunately, training their model requires expensive demonstrations. Therefore, we propose a training framework merely using a random agent in virtual environments. While the model trained in this manner exhibits good controls, it is limited by the random exploration possibilities. To address this limitation, we propose AutoExplore Agent - an exploration agent that entirely relies on the uncertainty of the world model, delivering diverse data from which it can learn the best. Our agent is fully independent of environment-specific rewards and thus adapts easily to new environments. With this approach, the pretrained multi-environment model can quickly adapt to new environments achieving video fidelity and controllability improvement.In order to obtain automatically large-scale interaction datasets for pretraining, we group environments with similar behavior and controls. To this end, we annotate the behavior and controls of 974 virtual environments - a dataset that we name RetroAct. For building our model, we first create an open implementation of Genie - GenieRedux and apply enhancements and adaptations in our version GenieRedux-G. Our code and data are available at https://github.com/insait-institute/GenieRedux.

We tackle the ill-posed inverse rendering problem in 3D reconstruction with a Neural Radiance Field (NeRF) approach informed by Physics-Based Rendering (PBR) theory, named PBR-NeRF. Our method addresses a key limitation in most NeRF and 3D Gaussian Splatting approaches: they estimate view-dependent appearance without modeling scene materials and illumination. To address this limitation, we present an inverse rendering (IR) model capable of jointly estimating scene geometry, materials, and illumination. Our model builds upon recent NeRF-based IR approaches, but crucially introduces two novel physics-based priors that better constrain the IR estimation. Our priors are rigorously formulated as intuitive loss terms and achieve state-of-the-art material estimation without compromising novel view synthesis quality. Our method is easily adaptable to other inverse rendering and 3D reconstruction frameworks that require material estimation. We demonstrate the importance of extending current neural rendering approaches to fully model scene properties beyond geometry and view-dependent appearance. Code is publicly available at: https://github.com/s3anwu/pbrnerf.

Recognizing arbitrary or previously unseen categories is essential for comprehensive real-world 3D scene understanding. Currently, all existing methods rely on 2D or textual modalities during training or together at inference. This highlights the clear absence of a model capable of processing 3D data alone for learning semantics end-to-end, along with the necessary data to train such a model. Meanwhile, 3D Gaussian Splatting (3DGS) has emerged as the de facto standard for 3D scene representation across various vision tasks. However, effectively integrating semantic reasoning into 3DGS in a generalizable manner remains an open challenge. To address these limitations, we introduce SceneSplat, to our knowledge the first large-scale 3D indoor scene understanding approach that operates natively on 3DGS. Furthermore, we propose a self-supervised learning scheme that unlocks rich 3D feature learning from unlabeled scenes. To power the proposed methods, we introduce SceneSplat-7K, the first large-scale 3DGS dataset for indoor scenes, comprising 7916 scenes derived from seven established datasets, such as ScanNet and Matterport3D. Generating SceneSplat-7K required computational resources equivalent to 150 GPU days on an L4 GPU, enabling standardized benchmarking for 3DGS-based reasoning for indoor scenes. Our exhaustive experiments on SceneSplat-7K demonstrate the significant benefit of the proposed method over the established baselines.

Scaling up model size and training data has advanced foundation models for instance-level perception, achieving state-of-the-art in-domain and zero-shot performance across object detection and segmentation. However, their high computational cost limits adoption on resource-constrained platforms. We first examine the limitations of existing architectures in enabling efficient edge deployment without compromising performance. We then introduce MOBIUS, a family of foundation models for universal instance segmentation, designed for Pareto-optimal downscaling to support deployment across devices ranging from high-end accelerators to mobile hardware. To reduce training and inference demands, we propose: (i) a bottleneck pixel decoder for efficient multi-scale and multi-modal fusion, (ii) a language-guided uncertainty calibration loss for adaptive decoder pruning, and (iii) a streamlined, unified training strategy. Unlike efficient baselines that trade accuracy for reduced complexity, MOBIUS reduces pixel and transformer decoder FLOPs by up to 55% and 75%, respectively, while maintaining state-of-the-art performance in just a third of the training iterations. MOBIUS establishes a new benchmark for efficient segmentation on both high-performance computing platforms and mobile devices.

Low-light image enhancement (LLIE) aims to improve the visibility of images captured in poorly lit environments. Prevalent event-based solutions primarily utilize events triggered by motion, i.e., ''motion events'' to strengthen only the edge texture, while leaving the high dynamic range and excellent low-light responsiveness of event cameras largely unexplored. This paper instead opens a new avenue from the perspective of estimating the illumination using ''temporal-mapping'' events, i.e., by converting the timestamps of events triggered by a transmittance modulation into brightness values. The resulting fine-grained illumination cues facilitate a more effective decomposition and enhancement of the reflectance component in low-light images through the proposed Illumination-aided Reflectance Enhancement module. Furthermore, the degradation model of temporal-mapping events under low-light condition is investigated for realistic training data synthesizing. To address the lack of datasets under this regime, we construct a beam-splitter setup and collect EvLowLight dataset that includes images, temporal-mapping events, and motion events. Extensive experiments across 5 synthetic datasets and our real-world EvLowLight dataset substantiate that the devised pipeline, dubbed RetinEV, excels in producing well-illuminated, high dynamic range images, outperforming previous state-of-the-art event-based methods by up to 6.62 dB, while maintaining an efficient inference speed of 35.6 frame-per-second on a 640X480 image.

Multimodal data is known to be helpful for visual tracking by improving robustness to appearance variations. However, sensor synchronization challenges often compromise data availability, particularly in video settings where shortages can be temporal. Despite its importance, this area remains underexplored. In this paper, we present the first comprehensive study on tracker performance with temporally incomplete multimodal data. Unsurprisingly, under such a circumstance, existing trackers exhibit significant performance degradation, as their rigid architectures lack the adaptability needed to effectively handle missing modalities. To address these limitations, we propose a flexible framework for robust multimodal tracking. We venture that a tracker should dynamically activate computational units based on missing data rates. This is achieved through a novel Heterogeneous Mixture-of-Experts fusion mechanism with adaptive complexity, coupled with a video-level masking strategy that ensures both temporal consistency and spatial completeness which is critical for effective video tracking. Surprisingly, our model not only adapts to varying missing rates but also adjusts to scene complexity. Extensive experiments show that our model achieves SOTA performance across 9 benchmarks, excelling in both conventional complete and missing modality settings. The code and benchmark will be publicly available at https://github.com/supertyd/FlexTrack/tree/main.

Fusing and balancing multi-modal inputs from novel sensors for dense prediction tasks, particularly semantic segmentation, is critically important yet remains a significant challenge. One major limitation is the tendency of multi-modal frameworks to over-rely on easily learnable modalities, a phenomenon referred to as unimodal dominance or bias. This issue becomes especially problematic in real-world scenarios where the dominant modality may be unavailable, resulting in severe performance degradation. To this end, we apply a simple but effective plug-and-play regularization term based on functional entropy, which introduces no additional parameters or modules. This term is designed to intuitively balance the contribution of each visual modality to the segmentation results. Specifically, we leverage the log-Sobolev inequality to bound functional entropy using functional-Fisher-information. By maximizing the information contributed by each visual modality, our approach mitigates unimodal dominance and establishes a more balanced and robust segmentation framework. A multi-scale regularization module is proposed to apply our proposed plug-and-play term on high-level features and also segmentation predictions for more balanced multi-modal learning. Extensive experiments on three datasets demonstrate that our proposed method achieves superior performance, i.e., +13.94%, +3.25%, and +3.64%, without introducing any additional parameters.

Understanding drivers’ decision-making is crucial for road safety. Although predicting the ego-vehicle’s path is valuable for driver-assistance systems, existing methods mainly focus on external factors like other vehicles’ motions, often neglecting the driver’s attention and intent. To address this gap, we infer the ego-trajectory by integrating the driver’s gaze and the surrounding scene. We introduce RouteFormer, a novel multimodal ego-trajectory prediction network combining GPS data, environmental context, and the driver's field-of-view—comprising first-person video and gaze fixations. We also present the Path Complexity Index (PCI), a new metric for trajectory complexity that enables a more nuanced evaluation of challenging scenarios. To tackle data scarcity and enhance diversity, we introduce GEM, a comprehensive dataset of urban driving scenarios enriched with synchronized driver field-of-view and gaze data. Extensive evaluations on GEM and DR(eye)VE demonstrate that RouteFormer significantly outperforms state-of-the-art methods, achieving notable improvements in prediction accuracy across diverse conditions. Ablation studies reveal that incorporating driver field-of-view data yields significantly better average displacement error, especially in challenging scenarios with high PCI scores, underscoring the importance of modeling driver attention. All data and code are available at meakbiyik.github.io/routeformer.

Multiple object tracking in complex scenarios - such as coordinated dance performances, team sports, or dynamic animal groups - presents unique challenges. In these settings, objects frequently move in coordinated patterns, occlude each other, and exhibit long-term dependencies in their trajectories. However, it remains a key open research question on how to model long-range dependencies within tracklets, interdependencies among tracklets, and the associated temporal occlusions. To this end, we introduce Samba, a novel linear-time set-of-sequences model designed to jointly process multiple tracklets by synchronizing the multiple selective state-spaces used to model each tracklet. Samba autoregressively predicts the future track query for each sequence while maintaining synchronized long-term memory representations across tracklets. By integrating Samba into a tracking-by-propagation framework, we propose SambaMOTR, the first tracker effectively addressing the aforementioned issues, including long-range dependencies, tracklet interdependencies, and temporal occlusions. Additionally, we introduce an effective technique for dealing with uncertain observations (MaskObs) and an efficient training recipe to scale SambaMOTR to longer sequences. By modeling long-range dependencies and interactions among tracked objects, SambaMOTR implicitly learns to track objects accurately through occlusions without any hand-crafted heuristics. Our approach significantly surpasses prior state-of-the-art on the DanceTrack, BFT, and SportsMOT datasets.

Recent progress in large language models and access to large-scale robotic datasets has sparked a paradigm shift in robotics models transforming them into generalists able to adapt to various tasks, scenes, and robot modalities. A large step for the community are open Vision Language Action models which showcase strong performance in a wide variety of tasks. In this work, we study the visual generalization capabilities of three existing robotic foundation models, and propose a corresponding evaluation framework. Our study shows that the existing models do not exhibit robustness to visual out-of-domain scenarios. This is potentially caused by limited variations in the training data and/or catastrophic forgetting, leading to domain limitations in the vision foundation models. We further explore OpenVLA, which uses two pre-trained vision foundation models and is, therefore, expected to generalize to out-of-domain experiments. However, we showcase catastrophic forgetting by DINO-v2 in OpenVLA through its failure to fulfill the task of depth regression. To overcome the aforementioned issue of visual catastrophic forgetting, we propose a gradual backbone reversal approach founded on model merging. This enables OpenVLA - which requires the adaptation of the visual backbones during initial training - to regain its visual generalization ability. Regaining this capability enables our ReVLA model to improve over OpenVLA by a factor of 77% and 66% for grasping and lifting in visual OOD tasks. Comprehensive evaluations, episode rollouts and model weights are available on the ReVLA Page

Current methods to learn controllers for autonomous vehicles (AVs) focus on behavioural cloning. Being trained only on exact historic data, the resulting agents often generalize poorly to novel scenarios. Simulators provide the opportunity to go beyond offline datasets, but they are still treated as complicated black boxes, only used to update the global simulation state. As a result, these RL algorithms are slow, sample-inefficient, and prior-agnostic. In this work, we leverage a differentiable simulator and design an analytic policy gradients (APG) approach to training AV controllers on the large-scale Waymo Open Motion Dataset. Our proposed framework brings the differentiable simulator into an end-to-end training loop, where gradients of the environment dynamics serve as a useful prior to help the agent learn a more grounded policy. We combine this setup with a recurrent architecture that can efficiently propagate temporal information across long simulated trajectories. This APG method allows us to learn robust, accurate, and fast policies, while only requiring widely-available expert trajectories, instead of scarce expert actions. We compare to behavioural cloning and find significant improvements in performance and robustness to noise in the dynamics, as well as overall more intuitive human-like handling.

Cross-Domain Few-Shot Object Detection (CD-FSOD) aims to detect novel objects with only a handful of labeled samples from previously unseen domains. While data augmentation and generative methods have shown promise in few-shot learning, their effectiveness for CD-FSOD remains unclear due to the need for both visual realism and domain alignment. Existing strategies, such as copy-paste augmentation and text-to-image generation, often fail to preserve the correct object category or produce backgrounds coherent with the target domain, making them non-trivial to apply directly to CD-FSOD. To address these challenges, we propose Domain-RAG, a training-free, retrieval-guided compositional image generation framework tailored for CD-FSOD. Domain-RAG consists of three stages: domain-aware background retrieval, domain-guided background generation, and foreground-background composition. Specifically, the input image is first decomposed into foreground and background regions. We then retrieve semantically and stylistically similar images to guide a generative model in synthesizing a new background, conditioned on both the original and retrieved contexts. Finally, the preserved foreground is composed with the newly generated domain-aligned background to form the generated image. Without requiring any additional supervision or training, Domain-RAG produces high-quality, domain-consistent samples across diverse tasks, including CD-FSOD, remote sensing FSOD, and camouflaged FSOD. Extensive experiments show consistent improvements over strong baselines and establish new state-of-the-art results. Codes will be released upon acceptance.The source code and instructions are available at https://github.com/LiYu0524/Domain-RAG.

3D Gaussian Splatting (3DGS) serves as a highly performant and efficient encoding of scene geometry, appearance, and semantics. Moreover, grounding language in 3D scenes has proven to be an effective strategy for 3D scene understanding. Current Language Gaussian Splatting line of work fall into three main groups: (i) per-scene optimization-based, (ii) per-scene optimization-free, and (iii) generalizable approach. However, most of them are evaluated only on rendered 2D views of a handful of scenes and viewpoints close to the training views, limiting ability and insight into holistic 3D understanding. To address this gap, we propose the first large-scale benchmark that systematically assesses these three groups of methods directly in 3D space, evaluating on 1060 scenes across three indoor datasets and one outdoor dataset. Benchmark results demonstrate a clear advantage of the generalizable paradigm, particularly in relaxing the scene-specific limitation, enabling fast feed-forward inference on novel scenes, and achieving superior segmentation performance. We further introduce SceneSplat-49K -- a carefully curated 3DGS dataset comprising of around 49K diverse indoor and outdoor scenes trained from multiple sources, with which we demonstrate generalizable approach could harness strong data priors. Our codes, benchmark, and datasets are available.

We propose LangHOPS, the first Multimodal Large Language Model (MLLM)-based framework for open-vocabulary object–part instance segmentation. Given an image, LangHOPS can jointly detect and segment hierarchical object and part instances from open-vocabulary candidate categories. Unlike prior approaches that rely on heuristic or learnable visual grouping, our approach grounds object–part hierarchies in language space. It integrates the MLLM into the object-part parsing pipeline to leverage rich knowledge and reasoning capabilities, and link multi-granularity concepts within the hierarchies. We evaluate LangHOPS across multiple challenging scenarios, including in-domain and cross-dataset object-part instance segmentation, and zero-shot semantic segmentation. LangHOPS achieves state-of-the-art results, surpassing previous methods by 5.5% Average Precision(AP) (in-domain) and 4.8% (cross-dataset) on the PartImageNet dataset and by 2.5% mIOU on unseen object parts in ADE20K (zero-shot). Ablation studies further validate the effectiveness of the language-grounded hierarchy and MLLM-driven part query refinement strategy.

World models have recently gained prominence for action-conditioned visual prediction in complex environments. However, relying on only a few recent observations causes them to lose long-term context. Consequently, within a few steps, the generated scenes drift from what was previously observed, undermining temporal coherence. This limitation, common in state-of-the-art world models, which are diffusion-based, stems from the lack of a lasting environment state. To address this problem, we introduce StateSpaceDiffuser, where a diffusion model is enabled to perform long-context tasks by integrating features from a state-space model, representing the entire interaction history. This design restores long-term memory while preserving the high-fidelity synthesis of diffusion models. To rigorously measure temporal consistency, we develop an evaluation protocol that probes a model’s ability to reinstantiate seen content in extended rollouts. Comprehensive experiments show that StateSpaceDiffuser significantly outperforms a strong diffusion-only baseline, maintaining a coherent visual context for an order of magnitude more steps. It delivers consistent views in both a 2D maze navigation and a complex 3D environment. These results establish that bringing state-space representations into diffusion models is highly effective in demonstrating both visual details and long-term memory. Project page: https://insait-institute.github.io/StateSpaceDiffuser/

3D visual grounding consists of identifying the instance in a 3D scene which is referred to by an accompanying language description. While several architectures have been proposed within the commonly employed grounding-by-selection framework, the utilized losses are comparatively under-explored. In particular, most methods rely on a basic supervised cross-entropy loss on the predicted distri-bution over candidate instances, which fails to model both spatial relations between instances and the internal fine-grained word-level structure of the verbal referral. Sparse attempts to additionally supervise verbal embeddings glob-ally by learning the class of the referred instance from the description or employing verbo- visual contrast to better separate instance embeddings do not fundamentally lift the aforementioned limitations. Responding to these shortcomings, we introduce two novel losses for 3D visual grounding: a visual-level offset loss on regressed vector offsets from each instance to the ground-truth referred instance and a language-related span loss on predictions for the word-level span of the referred instance in the description. In addition, we equip the verbo-visual fusion module of our new 3D visual grounding architecture AsphaltNet with a top-down bidirectional attentive fusion block, which en-ables the supervisory signals from our two losses to propa-gate to the respective converse branches of the network and thus aid the latter to learn context-aware instance embed-dings and grounding-aware verbal embeddings. AsphaltNet proposes novel auxiliary losses to aid 3D visual grounding with competitive results compared to the state-of-the-art on the ReferIt3D benchmark.

The increasing relevance of panoptic segmentation is tied to the advancements in autonomous driving and AR/VR applications. However, the deployment of such models has been limited due to the expensive nature of dense data an-notation, giving rise to unsupervised domain adaptation (UDA). A key challenge in panoptic UDA is reducing the domain gap between a labeled source and an unlabeled target domain while harmonizing the subtasks of seman-tic and instance segmentation to limit catastrophic inter-ference. While considerable progress has been achieved, existing approaches mainly focus on the adaptation of se-mantic segmentation. In this work, we focus on incorpo-rating instance-level adaptation via a novel instance-aware cross-domain mixing strategy IMix. IMix significantly en-hances the panoptic quality by improving instance segmen-tation performance. Specifically, we propose inserting high-confidence predicted instances from the target domain onto source images, retaining the exhaustiveness of the resulting pseudo-labels while reducing the injected confirmation bias. Nevertheless, such an enhancement comes at the cost of degraded semantic performance, attributed to catastrophic forgetting. To mitigate this issue, we regu-larize our semantic branch by employing CLIP-based do-main alignment (CDA), exploiting the domain-robustness of natural language prompts. Finally, we present an end-to-end model incorporating these two mechanisms called LIDAPS, achieving state-of-the-art results on all popular panoptic UDA benchmarks. https://github.com/elhamAm/LIDAPS

The Bird-Eye-View (BEV) is one of the most widely-used scene representations for visual perception in Autonomous Vehicles (AVs) due to its well suited compatibility to down-stream tasks. For the enhanced safety of AVs, modeling perception uncertainty in BEV is crucial. Recent diffusion-based methods offer a promising approach to uncertainty modeling for visual perception but fail to effectively detect small objects in the large coverage of the BEV. Such degradation of performance can be attributed primarily to the specific network architectures and the matching strategy used when training. Here, we address this problem by combining the diffusion paradigm with current state-of-the-art 3D object detectors in BEV. We analyze the unique challenges of this approach, which do not exist with deterministic detectors, and present a simple technique based on object query interpolation that allows the model to learn positional dependencies even in the presence of the diffusion noise. Based on this, we present a diffusion-based DETR model for object detection that bears similarities to particle methods. Abundant experimentation on the NuScenes dataset shows equal or better performance for our generative approach, compared to deterministic state-of-the-art methods. The source code is at https://github.com/insait-institute/ParticleDETR.

Nighttime scenes are hard to semantically perceive with learned models and annotate for humans. Thus, realistic synthetic nighttime data become all the more important for learning robust semantic perception at night, thanks to their accurate and cheap semantic annotations. However, existing data-driven or hand-crafted techniques for generating nighttime images from daytime counterparts suffer from poor realism. The reason is the complex interaction of highly spatially varying nighttime illumination, which differs drastically from its daytime counterpart, with objects of spatially varying materials in the scene, happening in 3D and being very hard to capture with such 2D approaches. The above 3D interaction and illumination shift have proven equally hard to model in the literature, as opposed to other conditions such as fog or rain. Our method, named Sun Off, Lights On (SOLO), is the first to perform nighttime simulation on single images in a photorealistic fashion by operating in 3D. It first explicitly estimates the 3D geometry, the materials and the locations of light sources of the scene from the input daytime image and relights the scene by probabilistically instantiating light sources in a way that accounts for their semantics and then running standard ray tracing. Not only is the visual quality and photorealism of our nighttime images superior to competing approaches including diffusion models, but the former images are also proven more beneficial for semantic nighttime segmentation in day-to-night adaptation. Code and data are publicly available at https://github.com/ktzevel/SOLO.

Abstract Neural implicit fields have recently shown increasing success in representing, learning and analysis of 3D shapes. Signed distance fields and occupancy fields are still the preferred choice of implicit representations with well-studied properties, despite their restriction to closed surfaces. With neural networks, unsigned distance fields as well as several other variations and training principles have been proposed with the goal to represent all classes of shapes. In this paper, we develop a novel and yet a fundamental representation considering unit vectors in 3D space and call it Vector Field (VF). At each point in $$\mathbb {R}^3$$ R 3 , VF is directed to the closest point on the surface. We theoretically demonstrate that VF can be easily transformed to surface density by computing the flux density. Unlike other standard representations, VF directly encodes an important physical property of the surface, its normal. We further show the advantages of VF representation, in learning open, closed, or multi-layered surfaces. We show that, thanks to the continuity property of the neural optimization with VF, a separate distance field becomes unnecessary for extracting surfaces from the implicit field via Marching Cubes. We compare our method on several datasets including ShapeNet where the proposed new neural implicit field shows superior accuracy in representing any type of shape, outperforming other standard methods. Codes are available at https://github.com/edomel/ImplicitVF .

The progress on Hyperspectral image (HSI) super-resolution (SR) is still lagging behind the research of RGB image SR. HSIs usually have a high number of spectral bands, so accurately modeling spectral band interaction for HSI SR is hard. Also, training data for HSI SR is hard to obtain so the dataset is usually rather small. In this work, we propose a new test-time training method to tackle this problem. Specifically, a novel self-training framework is developed, where more accurate pseudo-labels and more accurate LR-HR relationships are generated so that the model can be further trained with them to improve performance. In order to better support our test-time training method, we also propose a new network architecture to learn HSI SR without modeling spectral band interaction and propose a new data augmentation method Spectral Mixup to increase the diversity of the training data at test time. We also collect a new HSI dataset with a diverse set of images of interesting objects ranging from food to vegetation, to materials, and to general scenes. Extensive experiments on multiple datasets show that our method can improve the performance of pre-trained models significantly after test-time training and outperform competing methods significantly for HSI SR.

Adaptation of semantic segmentation networks to different visual conditions is vital for robust perception in autonomous cars and robots. However, previous work has shown that most feature-level adaptation methods, which employ adversarial training and are validated on synthetic-to-real adaptation, provide marginal gains in condition-level adaptation, being outperformed by simple pixel-level adaptation via stylization. Motivated by these findings, we propose to leverage stylization in performing feature-level adaptation by aligning the internal network features extracted by the encoder of the network from the original and the stylized view of each input image with a novel feature invariance loss. In this way, we encourage the encoder to extract features that are already invariant to the style of the input, allowing the decoder to focus on parsing these features and not on further abstracting from the specific style of the input. We implement our method, named Condition-Invariant Semantic Segmentation (CISS), on the current state-of-the-art domain adaptation architecture and achieve outstanding results on condition-level adaptation. In particular, CISS sets the new state of the art in the popular daytime-to-nighttime Cityscapes Dark Zurich benchmark. Furthermore, our method achieves the second-best performance on the normal-to-adverse Cityscapes ACDC benchmark. CISS is shown to generalize well to domains unseen during training, such as BDD100K-night and ACDC-night.

Effective video frame interpolation hinges on the adept handling of motion in the input scene. Prior work acknowledges asynchronous event information for this, but often overlooks whether motion induces blur in the video, limiting its scope to sharp frame interpolation. We instead propose a unified framework for event-based frame interpolation that performs deblurring ad-hoc and thus works both on sharp and blurry input videos. Our model consists in a bidirectional recurrent network that incorporates the temporal dimension of interpolation and fuses information from the input frames and the events adaptively based on their temporal proximity. To enhance the generalization from synthetic data to real event cameras, we integrate self-supervised framework with the proposed model to enhance the generalization on real-world datasets in the wild. At the dataset level, we introduce a novel real-world high-resolution dataset with events and color videos named HighREV, which provides a challenging evaluation setting for the examined task. Extensive experiments show that our network consistently outperforms previous state-of-the-art methods on frame interpolation, single image deblurring, and the joint task of both. Experiments on domain transfer reveal that self-supervised training effectively mitigates the performance degradation observed when transitioning from synthetic data to real-world data. Code and datasets are available at https://github.com/AHupuJR/REFID.

In this paper, we propose a novel cross-attention-based generative adversarial network (GAN) for the challenging person image generation task. Cross-attention is a novel and intuitive multi-modal fusion method in which an attention/correlation matrix is calculated between two feature maps of different modalities. Specifically, we propose the novel XingGAN (or CrossingGAN), which consists of two generation branches that capture the person's appearance and shape, respectively. Moreover, we propose two novel cross-attention blocks to effectively transfer and update the person's shape and appearance embeddings for mutual improvement. This has not been considered by any other existing GAN-based image generation work. To further learn the long-range correlations between different person poses at different scales and sub-regions, we propose two novel multi-scale cross-attention blocks. To tackle the issue of independent correlation computations within the cross-attention mechanism leading to noisy and ambiguous attention weights, which hinder performance improvements, we propose a module called enhanced attention (EA). Lastly, we introduce a novel densely connected co-attention module to fuse appearance and shape features at different stages effectively. Extensive experiments on two public datasets demonstrate that the proposed method outperforms current GAN-based methods and performs on par with diffusion-based methods. However, our method is significantly faster than diffusion-based methods in both training and inference.

We propose a novel spatial-temporal graph Mamba (STG-Mamba) for the music-guided dance video synthesis task, i.e., to translate the input music to a dance video. STG-Mamba consists of two translation mappings: music-to-skeleton translation and skeleton-to-video translation. In the music-to-skeleton translation, we introduce a novel spatial-temporal graph Mamba (STGM) block to effectively construct skeleton sequences from the input music, capturing dependencies between joints in both the spatial and temporal dimensions. For the skeleton-to-video translation, we propose a novel self-supervised regularization network to translate the generated skeletons, along with a conditional image, into a dance video. Lastly, we collect a new skeleton-to-video translation dataset from the Internet, containing 54,944 video clips. Extensive experiments demonstrate that STG-Mamba achieves significantly better results than existing methods.

The Diffusion Model (DM) has emerged as the SOTA approach for image synthesis. However, the existing DM cannot perform well on some image-to-image translation (I2I) tasks. Different from image synthesis, some I2I tasks, such as super-resolution, require generating results in accordance with GT images. Traditional DMs for image synthesis require extensive iterations and large denoising models to estimate entire images, which gives their strong generative ability but also leads to artifacts and inefficiency for I2I. To tackle this challenge, we propose a simple, efficient, and powerful DM framework for I2I, called DiffI2I. Specifically, DiffI2I comprises three key components: a compact I2I prior extraction network (CPEN), a dynamic I2I transformer (DI2Iformer), and a denoising network. We train DiffI2I in two stages: pretraining and DM training. For pretraining, GT and input images are fed into CPEN to capture a compact I2I prior representation (IPR) guiding DI2Iformer. In the second stage, the DM is trained to only use the input images to estimate the same IRP as CPEN. Compared to traditional DMs, the compact IPR enables DiffI2I to obtain more accurate outcomes and employ a lighter denoising network and fewer iterations. Through extensive experiments on various I2I tasks, we demonstrate that DiffI2I achieves SOTA performance while significantly reducing computational burdens.

This paper revisits few-shot 3D point cloud semantic segmentation (FS-PCS), with a focus on two significant is-sues in the state-of-the-art: foreground leakage and sparse point distribution. The former arises from non-uniform point sampling, allowing models to distinguish the density disparities between foreground and background for easier segmentation. The latter results from sampling only 2,048 points, limiting semantic information and deviating from the real-world practice. To address these issues, we in-troduce a standardized FS-PCS setting, upon which a new benchmark is built. Moreover, we propose a novel FS-PCS model. While previous methods are based on feature op-timization by mainly refining support features to enhance prototypes, our method is based on correlation optimization, referred to as Correlation Optimization Segmentation (COSeg). Specifically, we compute Class-specific Multi-prototypical Correlation (CMC) for each query point, rep-resenting its correlations to category prototypes. Then, we propose the Hyper Correlation Augmentation (HCA) mod-ule to enhance CMC. Furthermore, tackling the inherent property of few-shot training to incur base susceptibility for models, we propose to learn non-parametric prototypes for the base classes during training. The learned base proto-types are used to calibrate correlations for the background class through a Base Prototypes Calibration (BPC) module. Experiments on popular datasets demonstrate the superior-ity of COSeg over existing methods. The code is available at github.com/ZhaochongAnICOSeg.

Diffusion model-based image restoration (IR) aims to use diffusion models to recover high-quality (HQ) images from degraded images, achieving promising performance. Due to the inherent property of diffusion models, most existing methods need long serial sampling chains to restore HQ images step-by-step, resulting in expensive sampling time and high computation costs. Moreover, such long sampling chains hinder understanding the relationship between inputs and restoration results since it is hard to compute the gra-dients in the whole chains. In this work, we aim to rethink the diffusion model-based IR models through a different per-spective, i.e., a deep equilibrium (DEQ) fixed point system, called DeqIR. Specifically, we derive an analytical solution by modeling the entire sampling chain in these IR models as a joint multivariate fixed point system. Based on the analyti-cal solution, we can conduct parallel sampling and restore HQ images without training. Furthermore, we compute fast gradients via DEQ inversion and found that initialization optimization can boost image quality and control the gen-eration direction. Extensive experiments on benchmarks demonstrate the effectiveness of our method on typical IR tasks and real-world settings.

Extracting keypoint locations from input hand frames, known as 3D hand pose estimation, is a critical task in various human-computer interaction applications. Essentially, the 3D hand pose estimation can be regarded as a 3D point subset generative problem conditioned on input frames. Thanks to the recent significant progress on diffusion-based generative models, hand pose estimation can also benefit from the diffusion model to estimate keypoint locations with high quality. However, directly deploying the existing diffusion models to solve hand pose estimation is non-trivial, since they cannot achieve the complex permutation mapping and precise localization. Based on this motivation, this paper proposes HandDiff, a diffusion-based hand pose estimation model that iteratively denoises accurate hand pose conditioned on hand-shaped image-point clouds. In order to recover keypoint permutation and accurate location, we further introduce Joint-wise condition and local detail condition. Experimental results demonstrate that the proposed HandDiff significantly outperforms the existing approaches on four challenging hand pose benchmark datasets. Codes and pre-trained models are publicly available at https://github.com/cwc1260/HandDiff.

The recently increased role of mobile photography has raised the standards of on-device photo processing tremendously. Despite the latest advancements in camera hardware, the mobile camera sensor area cannot be increased significantly due to physical constraints, leading to a pixel size of 0.6-2.0μm, which results in strong image noise even in moderate lighting conditions. In the era of deep learning, one can train a CNN model to perform robust image denoising. However, there is still a lack of a substantially diverse dataset for this task. To address this problem, we introduce a novel Mobile Image Denoising Dataset (MIDD) comprising over 400,000 noisy / noise-free image pairs captured under various conditions by 20 different mobile camera sensors. Additionally, we propose a new DPreview test set consisting of data from 294 different cameras for precise model evaluation. Furthermore, we present the efficient baseline model SplitterNet for the considered mobile image denoising task that achieves high numerical and visual results, while being able to process 8MP photos directly on smartphone GPUs in under one second. Thereby outperforming models with similar runtimes. This model is also compatible with recent mobile NPUs, demonstrating an even higher speed when deployed on them. The conducted experiments demonstrate high robustness of the proposed solution when applied to images from previously unseen sensors, showing its high generalizability. The datasets, code and models can be found on the official project website11https://people.ee.ethz.ch/-ihnatova/midd.html22https://github.com/rflepp/Efficient_Mobile_Denoising_Models.

The estimation of implicit cross-frame correspondences and the high computational cost have long been major chal-lenges in video semantic segmentation (VSS) for driving scenes. Prior works utilize keyframes, feature propagation, or cross-frame attention to address these issues. By contrast, we are the first to harness vanishing point (VP) priors for more effective segmentation. Intuitively, objects near VPs (i.e., away from the vehicle) are less discernible. Moreover, they tend to move radially away from the VP over time in the usual case of a forward-facing camera, a straight road, and linear forward motion of the vehicle. Our novel, efficient network for VSS, named VPSeg, incor-porates two modules that utilize exactly this pair of static and dynamic VP priors: sparse-to-dense feature mining (DenseVP) and VP-guided motion fusion (MotionVP). MotionVP employs VP-guided motion estimation to establish explicit correspondences across frames and help attend to the most relevant features from neighboring frames, while Dense Vp enhances weak dynamic features in distant re-gions around VPs. These modules operate within a context-detail framework, which separates contextual features from high-resolution local features at different input resolutions to reduce computational costs. Contextual and local fea-tures are integrated through contextualized motion attention (CMA) for the final prediction. Extensive experiments on two popular driving segmentation benchmarks, Cityscapes and ACDC, demonstrate that VPSeg outperforms previous SOTA methods, with only modest computational overhead. The resources are available at https://github.com/RascalGdd/VPSeg.

Recovering the 3D scene geometry from a single view is a fundamental yet ill-posed problem in computer vision. While classical depth estimation methods infer only a 2.5D scene representation limited to the image plane, recent approaches based on radiance fields reconstruct a full 3D representation. However, these methods still struggle with occluded regions since inferring geometry without visual observation requires (i) semantic knowledge of the surroundings, and (ii) reasoning about spatial context. We propose KYN, a novel method for single-view scene reconstruction that reasons about semantic and spatial context to predict each point's density. We introduce a vision-language modulation module to enrich point features with fine-grained semantic information. We aggregate point representations across the scene through a language-guided spatial attention mechanism to yield per-point density predictions aware of the 3D semantic context. We show that KYN improves 3D shape recovery compared to predicting density for each 3D point in isolation. We achieve state-of-the-art results in scene and object reconstruction on KITTI-360, and show improved zero-shot generalization compared to prior work. Project page: https://ruili3.github.io/kyn.

The robust association of the same objects across video frames in complex scenes is crucial for many applications, especially multiple object tracking (MOT). Current methods predominantly rely on labeled domain-specific video datasets, which limits the cross-domain generalization of learned similarity embeddings. We propose MASA, a novel method for robust instance association learning, capable of matching any objects within videos across diverse domains without tracking labels. Leveraging the rich object segmentation from the Segment Anything Model (SAM), MASA learns instance-level correspondence through exhaustive data transformations. We treat the SAM outputs as dense object region proposals and learn to match those regions from a vast image collection. We further design a universal MASA adapter which can work in tandem with foundational segmentation or detection models and enable them to track any detected objects. Those combinations present strong zero-shot tracking ability in complex domains. Extensive tests on multiple challenging MOT and MOTS benchmarks indicate that the proposed method, using only unlabeled static images, achieves even better performance than state-of-the-art methods trained with fully annotated in-domain video sequences, in zero-shot association. Our code is available at github.com/siyuanliii/masa.

Neural RGBD SLAM techniques have shown promise in dense Simultaneous Localization And Mapping (SLAM), yet face challenges such as error accumulation during cam-era tracking resulting in distorted maps. In response, we introduce Loopy-SLAM that globally optimizes poses and the dense 3D model. We use frame-to-model tracking using a data-driven point-based submap generation method and trigger loop closures online by performing global place recognition. Robust pose graph optimization is used to rigidly align the local submaps. As our representation is point based, map corrections can be performed efficiently without the need to store the entire history of input frames used for mapping as typically required by methods employing a grid based mapping structure. Evaluation on the synthetic Replica and real-world TUM-RGBD and Scan-Net datasets demonstrate competitive or superior performance in tracking, mapping, and rendering accuracy when compared to existing dense neural RGBD SLAM methods. Project page: notchla. gi thub. io/Loopy-SLAM/.

In this paper, we showcase the effectiveness of optimizing monocular camera poses as a continuous function of time. The camera poses are represented using an implicit neural function which maps the given time to the corresponding camera pose. The mapped camera poses are then used for the downstream tasks where joint camera pose optimization is also required. While doing so, the network parameters - that implicitly represent camera poses - are optimized. We exploit the proposed method in four diverse experimental settings, namely, (1) NeRF from noisy poses; (2) NeRF from asynchronous Events; (3) Visual Simultaneous Localization and Mapping (vSLAM); and (4) vSLAM with IMUs. In all four settings, the proposed method performs significantly better than the compared baselines and the state-of-the-art methods. Additionally, using the assumption of continuous motion, changes in pose may actually live in a manifold that has lower than 6 degrees of freedom (DOF) is realized. We call this low DOF motion representation as the intrinsic motion and use the approach in vSLAM settings, showing impressive camera tracking performance. We release our code at: https://github.com/qimaqi/Continuous-Pose-in-NeRF.

We study object interaction anticipation in egocentric videos. This task requires an understanding of the spatio-temporal context formed by past actions on objects, coined action context. We propose TransFusion, a multimodal transformer-based architecture for short-term object interaction anticipation. Our method exploits the representational power of language by summarizing the action con-text textually, after leveraging pre-trained vision-language foundation models to extract the action context from past video frames. The summarized action context and the last observed video frame are processed by the multimodal fusion module to forecast the next object interaction. Experiments on the Ego4D next active object interaction dataset show the effectiveness of our multimodal fusion model and highlight the benefits of using the power of foundation models and language-based context summaries in a task where vision may appear to suffice. Our novel approach outperforms all state-of-the-art methods on both versions of the Ego4D dataset. A project video and code are available at https://eth-ait.github.io/transfusion-proj/.

Canonical emotions, such as happy, sad, and fearful, are easy to understand and annotate. However, emotions are of-ten compound, e.g. happily surprised, and can be mapped to the action units (AUs) used for expressing emotions, and trivially to the canonical ones. Intuitively, emotions are continuous as represented by the arousal-valence (AV) model. An interpretable unification of these four modalities -namely, Canonical, Compound, AUs, and AV- is highly desirable, for a better representation and understanding of emotions. However, such unification remains to be unknown in the current literature. In this work, we propose an in-terpretable and unified emotion model, referred as C2A2. We also develop a method that leverages labels of the non-unified models to annotate the novel unified one. Finally, we modify the text-conditional diffusion models to under-stand continuous numbers, which are then used to generate continuous expressions using our unified emotion model. Through quantitative and qualitative experiments, we show that our generated images are rich and capture subtle ex-pressions. Our work allows a fine- grained generation of ex-pressions in conjunction with other textual inputs and offers a new label space for emotions at the same time.

Accurate monocular metric depth estimation (MMDE) is crucial to solving downstream tasks in 3D perception and modeling. However, the remarkable accuracy of recent MMDE methods is confined to their training domains. These methods fail to generalize to unseen domains even in the presence of moderate domain gaps, which hinders their practical applicability. We propose a new model, UniDepth, capable of reconstructing metric 3D scenes from solely single images across domains. Departing from the existing MMDE methods, UniDepth directly predicts metric 3D points from the input image at inference time without any additional information, striving for a universal and flexible MMDE solution. In particular, UniDepth implements a self-promptable camera module predicting dense camera representation to condition depth features. Our model exploits a pseudo-spherical output representation, which disentangles camera and depth representations. In addition, we propose a geometric invariance loss that promotes the invariance of camera-prompted depth features. Thorough evaluations on ten datasets in a zero-shot regime consistently demonstrate the superior performance of UniDepth, even when compared with methods directly trained on the testing domains. Code and models are available at: github.com/lpiccinelli-eth/unidepth.

In the realm of video object tracking, auxiliary modalities such as depth, thermal, or event data have emerged as valuable assets to complement the RGB trackers. In practice, most existing RGB trackers learn a single set of parameters to use them across datasets and applications. However, a similar single-model unification for multi-modality tracking presents several challenges. These challenges stem from the inherent heterogeneity of inputs - each with modality-specific representations, the scarcity of multi-modal datasets, and the absence of all the modalities at all times. In this work, we introduce Un-Track, a Unified Tracker of a single set of parameters for any modality. To handle any modality, our method learns their common latent space through low-rank factorization and reconstruction techniques. More importantly, we use only the RGB-X pairs to learn the common latent space. This unique shared representation seamlessly binds all modalities together, enabling effective unification and accommodating any missing modality, all within a single transformer-based architecture. Our Un-Track achieves +8.1 absolute F-score gain, on the DepthTrack dataset, by introducing only +2.14 (over 21.50) GFLOPs with +6.6M (over 93M) parameters, through a simple yet efficient prompting strategy. Extensive comparisons on five benchmark datasets with different modalities show that Un-Track surpasses both SOTA unified trackers and modality-specific counterparts, validating our effectiveness and practicality. The source code is publicly available at https://thub.com/Zongwei97/UnTrack.

Adiabatic quantum computing (AQC) is a promising approach for discrete and often NP-hard optimization prob-lems. Current AQCs allow to implement problems of re-search interest, which has sparked the development of quan-tum representations for many computer vision tasks. De-spite requiring multiple measurements from the noisy AQC, current approaches only utilize the best measurement, dis-carding information contained in the remaining ones. In this work, we explore the potential of using this information for probabilistic balanced k-means clustering. Instead of discarding non-optimal solutions, we propose to use them to compute calibrated posterior probabilities with little ad-ditional compute cost. This allows us to identify ambiguous solutions and data points, which we demonstrate on a D-Wave AQC on synthetic tasks and real visual data.

Multi-modality image fusion is a technique that combines information from different sensors or modalities, en-abling the fused image to retain complementary features from each modality, such as functional highlights and texture details. However, effective training of such fusion models is challenging due to the scarcity of ground truth fusion data. To tackle this issue, we propose the Equivariant Multi-Modality imAge fusion (EMMA) paradigm for end-to-end self-supervised learning. Our approach is rooted in the prior knowledge that natural imaging responses are equiv-ariant to certain transformations. Consequently, we introduce a novel training paradigm that encompasses a fusion module, a pseudo-sensing module, and an equivariant fusion module. These components enable the net training to follow the principles of the natural sensing-imaging process while satisfying the equivariant imaging prior. Extensive experiments confirm that EMMA yields high-quality fusion results for infraredvisible and medical images, concurrently facilitating downstream multi-modal segmentation and detection tasks. The code is available at https://github.com/Zhaozixiang1228/MMIF-EMMA.

We introduce Stereo Risk, a new deep-learning approach to solve the classical stereo-matching problem in computer vision. As it is well-known that stereo matching boils down to a per-pixel disparity estimation problem, the popular state-of-the-art stereo-matching approaches widely rely on regressing the scene disparity values, yet via discretization of scene disparity values. Such discretization often fails to capture the nuanced, continuous nature of scene depth. Stereo Risk departs from the conventional discretization approach by formulating the scene disparity as an optimal solution to a continuous risk minimization problem, hence the name "stereo risk". We demonstrate that $L^1$ minimization of the proposed continuous risk function enhances stereo-matching performance for deep networks, particularly for disparities with multi-modal probability distributions. Furthermore, to enable the end-to-end network training of the non-differentiable $L^1$ risk optimization, we exploited the implicit function theorem, ensuring a fully differentiable network. A comprehensive analysis demonstrates our method's theoretical soundness and superior performance over the state-of-the-art methods across various benchmark datasets, including KITTI 2012, KITTI 2015, ETH3D, SceneFlow, and Middlebury 2014.

Lightweight image super-resolution (SR) methods have obtained promising results with moderate model complexity. These approaches primarily focus on a lightweight architecture design, but neglect to further reduce network redundancy. While some model compression techniques try to achieve more lightweight SR models with neural architecture search, knowledge distillation, or channel pruning, they typically require considerable extra computational resources or neglect to prune weights. To address these issues, we propose a flexible meta pruning (FMP) for lightweight image SR, where the network channels and weights are pruned simultaneously. Specifically, we control the network sparsity via channel vectors and weight indicators. We feed them into a hypernetwork, whose parameters act as meta-data for the parameters of the SR backbone. Consequently, for each network layer, we conduct structured pruning with channel vectors, which control the output and input channels. Besides, we conduct unstructured pruning with weight indicators to influence the sparsity of kernel weights, resulting in flexible pruning. During pruning, the sparsity of both channel vectors and weight indicators are regularized. We optimize the channel vectors and weight indicators with proximal gradient and SGD. We conduct extensive experiments to investigate critical factors in the flexible channel and weight pruning for image SR, demonstrating the superiority of our FMP when applied to baseline image SR architectures.

Image fusion integrates essential information from multiple images into a single composite, enhancing structures, textures, and refining imperfections. Existing methods predominantly focus on pixel-level and semantic visual features for recognition, but often overlook the deeper text-level semantic information beyond vision. Therefore, we introduce a novel fusion paradigm named image Fusion via vIsion-Language Model (FILM), for the first time, utilizing explicit textual information from source images to guide the fusion process. Specifically, FILM generates semantic prompts from images and inputs them into ChatGPT for comprehensive textual descriptions. These descriptions are fused within the textual domain and guide the visual information fusion, enhancing feature extraction and contextual understanding, directed by textual semantic information via cross-attention. FILM has shown promising results in four image fusion tasks: infrared-visible, medical, multi-exposure, and multi-focus image fusion. We also propose a vision-language dataset containing ChatGPT-generated paragraph descriptions for the eight image fusion datasets across four fusion tasks, facilitating future research in vision-language model-based image fusion. Code and dataset are available at https://github.com/Zhaozixiang1228/IF-FILM.

In this work, we address the challenge of deploying Neural Radiance Field (NeRFs) in Simultaneous Localization and Mapping (SLAM) under the condition of lacking depth information, relying solely on RGB inputs. The key to unlocking the full potential of NeRF in such a challenging context lies in the integration of real-world priors. A crucial prior we explore is the binary opacity prior of 3D space with opaque objects. To effectively incorporate this prior into the NeRF framework, we introduce a ternary-type opacity (TT) model instead, which categorizes points on a ray intersecting a surface into three regions: before, on, and behind the surface. This enables a more accurate rendering of depth, subsequently improving the performance of image warping techniques. Therefore, we further propose a novel hybrid odometry (HO) scheme that merges bundle adjustment and warping-based localization. Our integrated approach of TT and HO achieves state-of-the-art performance on synthetic and real-world datasets, in terms of both speed and accuracy. This breakthrough underscores the potential of NeRF-SLAM in navigating complex environments with high fidelity.

Moving object segmentation (MOS) in dynamic scenes is an important, challenging, but under-explored research topic for autonomous driving, especially for sequences obtained from moving ego vehicles. Most segmentation methods leverage motion cues obtained from optical flow maps. However, since these methods are often based on optical flows that are pre-computed from successive RGB frames, this neglects the temporal consideration of events occurring within the inter-frame, consequently constraining its ability to discern objects exhibiting relative staticity but genuinely in motion. To address these limitations, we propose to exploit event cameras for better video understanding, which provide rich motion cues without relying on optical flow. To foster research in this area, we first introduce a novel large-scale dataset called DSEC-MOS for moving object segmentation from moving ego vehicles, which is the first of its kind. For benchmarking, we select various mainstream methods and rigorously evaluate them on our dataset. Subsequently, we devise EmoFormer, a novel network able to exploit the event data. For this purpose, we fuse the event temporal prior with spatial semantic maps to distinguish genuinely moving objects from the static background, adding another level of dense supervision around our object of interest. Our proposed network relies only on event data for training but does not require event input during inference, making it directly comparable to frame-only methods in terms of efficiency and more widely usable in many application cases. The exhaustive comparison highlights a significant performance improvement of our method over all other methods. The source code and dataset are publicly available at: https://github.com/ZZYZhou/DSEC-MOS.

Neural implicit functions have demonstrated significant importance in various areas such as computer vision, graphics. Their advantages include the ability to represent complex shapes and scenes with high fidelity, smooth interpolation capabilities, and continuous representations. Despite these benefits, the development and analysis of implicit functions have been limited by the lack of comprehensive datasets and the substantial computational resources required for their implementation and evaluation. To address these challenges, we introduce "Implicit-Zoo": a large-scale dataset requiring thousands of GPU training days designed to facilitate research and development in this field. Our dataset includes diverse 2D and 3D scenes, such as CIFAR-10, ImageNet-1K, and Cityscapes for 2D image tasks, and the OmniObject3D dataset for 3D vision tasks. We ensure high quality through strict checks, refining or filtering out low-quality data. Using Implicit-Zoo, we showcase two immediate benefits as it enables to: (1) learn token locations for transformer models; (2) Directly regress 3D cameras poses of 2D images with respect to NeRF models. This in turn leads to an \emph{improved performance} in all three task of image classification, semantic segmentation, and 3D pose regression -- thereby unlocking new avenues for research.

Image Restoration (IR), a classic low-level vision task, has witnessed significant advancements through deep models that effectively model global information. Notably, the emergence of Vision Transformers (ViTs) has further propelled these advancements. When computing, the self-attention mechanism, a cornerstone of ViTs, tends to encompass all global cues, even those from semantically unrelated objects or regions. This inclusivity introduces computational inefficiencies, particularly noticeable with high input resolution, as it requires processing irrelevant information, thereby impeding efficiency. Additionally, for IR, it is commonly noted that small segments of a degraded image, particularly those closely aligned semantically, provide particularly relevant information to aid in the restoration process, as they contribute essential contextual cues crucial for accurate reconstruction. To address these challenges, we propose boosting IR's performance by sharing the key semantics via Transformer for IR (i.e., SemanIR) in this paper. Specifically, SemanIR initially constructs a sparse yet comprehensive key-semantic dictionary within each transformer stage by establishing essential semantic connections for every degraded patch. Subsequently, this dictionary is shared across all subsequent transformer blocks within the same stage. This strategy optimizes attention calculation within each block by focusing exclusively on semantically related components stored in the key-semantic dictionary. As a result, attention calculation achieves linear computational complexity within each window. Extensive experiments across 6 IR tasks confirm the proposed SemanIR's state-of-the-art performance, quantitatively and qualitatively showcasing advancements. The visual results, code, and trained models are available at: https://github.com/Amazingren/SemanIR.

Generic Object Tracking (GOT) is the problem of tracking target objects, specified by bounding boxes in the first frame of a video. While the task has received much attention in the last decades, researchers have almost exclusively focused on the single object setting. However multiobject GOT poses its own challenges and is more attractive in real-world applications. We attribute the lack of research interest into this problem to the absence of suitable benchmarks. In this work, we introduce a new largescale GOT benchmark, LaGOT, containing multiple annotated target objects per sequence. Our benchmark allows users to tackle key remaining challenges in GOT, aiming to increase robustness and reduce computation through joint tracking of multiple objects simultaneously. In addition, we propose a transformer-based GOT tracker baseline capable of joint processing of multiple objects through shared computation. Our approach achieves a 4× faster run-time in case of 10 concurrent objects compared to tracking each object independently and outperforms existing single object trackers on our new benchmark. In addition, our approach achieves highly competitive results on single-object GOT datasets, setting a new state of the art on TrackingNet with a success rate AUC of 84.4%. Our benchmark, code, results and trained models are available at https://github.com/visionml/pytracking.

Monitoring a fleet of robots requires stable long-term tracking with re-identification, which is yet an unsolved challenge in many scenarios. One application of this is the analysis of autonomous robotic soccer games at RoboCup. Tracking in these games requires handling of identically looking players, strong occlusions, and non-professional video recordings, but also offers state information estimated by the robots. In order to make effective use of the information coming from the robot sensors, we propose a robust tracking and identification pipeline. It fuses external non-calibrated camera data with the robots’ internal states using quadratic optimization for tracklet matching. The approach is validated using game recordings from previous RoboCup World Cup tournaments.

Event-based sensing is a relatively new imaging modality that enables low latency, low power, high temporal resolution and high dynamic range acquisition. These properties make it a highly desirable sensor for edge applications and in high dynamic range environments. As of today, most event-based sensors are monochromatic (grayscale), capturing light from a wide spectral range over the visible, in a single channel. In this paper, we introduce multispectral events and study their advantages. In particular, we consider multiple bands in the visible and near-infrared range, and explore their potential compared to monochromatic events and conventional multispectral imaging for the face detection task. We further release the first large scale bimodal face detection datasets, with RGB videos and their simulated color events, N-MobiFace and N-YoutubeFaces, and a smaller dataset with multispectral videos and events, N-SpectralFace. We find that early fusion of multispectral events significantly improves the face detection performance, compared to the early fusion of conventional multi-spectral images. This result shows that multispectral events carry relatively more useful information about the scene than conventional multispectral images do, with respect to their grayscale equivalent. To the best of our knowledge, our proposed method is the first exploratory research on multispectral events, specifically including near infrared data.

We propose a discrete latent distribution for Generative Adversarial Networks (GANs). Instead of drawing latent vectors from a continuous prior, we sample from a finite set of learnable latents. However, a direct parametrization of such a distribution leads to an intractable linear increase in memory in order to ensure sufficient sample diversity. We address this key issue by taking inspiration from the encoding of information in biological organisms. Instead of learning a separate latent vector for each sample, we split the latent space into a set of genes. For each gene, we train a small bank of gene variants. Thus, by independently sampling a variant for each gene and combining them into the final latent vector, our approach can represent a vast number of unique latent samples from a compact set of learnable parameters. Interestingly, our gene-inspired latent encoding allows for new and intuitive approaches to latent-space exploration, enabling conditional sampling from our unconditionally trained model. Moreover, our approach preserves state-of-the-art photo-realism while achieving better disentanglement than the widely-used StyleMapping network.

The DEtection TRansformer (DETR) opened new possibilities for object detection by modeling it as a translation task: converting image features into object-level representations. Previous works typically add expensive modules to DETR to perform Multi-Object Tracking (MOT), resulting in more complicated architectures. We instead show how DETR can be turned into a MOT model by employing an instance-level contrastive loss, a revised sampling strategy and a lightweight assignment method. Our training scheme learns object appearances while preserving detection capabilities and with little overhead. Its performance surpasses the previous state-of-the-art by +2.6 mMOTA on the challenging BDD100K dataset and is comparable to existing transformer-based methods on the MOT17 dataset.

As 3D perception problems grow in popularity and the need for large-scale labeled datasets for LiDAR semantic segmentation increase, new methods arise that aim to reduce the necessity for dense annotations by employing weakly-supervised training. However these methods continue to show weak boundary estimation and high false negative rates for small objects and distant sparse regions. We argue that such weaknesses can be compensated by using RGB images which provide a denser representation of the scene. We propose an image-guidance network (IGNet) which builds upon the idea of distilling high level feature information from a domain adapted synthetically trained 2D semantic segmentation network. We further utilize a one-way contrastive learning scheme alongside a novel mixing strategy called FOVMix, to combat the horizontal field-of-view mismatch between the two sensors and enhance the effects of image guidance. IGNet achieves state-of-the-art results for weakly-supervised LiDAR semantic segmentation on ScribbleKITTI, boasting up to 98% relative performance to fully supervised training with only 8% labeled points, while introducing no additional annotation burden or computational/memory cost during inference. Furthermore, we show that our contributions also prove effective for semi-supervised training, where IGNet claims state-of-the-art results on both ScribbleKITTI and SemanticKITTI.

We introduce an improved solution to the neural image-based rendering problem in computer vision. Given a set of images taken from a freely moving camera at train time, the proposed approach could synthesize a realistic image of the scene from a novel viewpoint at test time. The key ideas presented in this paper are (i) Recovering accurate camera parameters via a robust pipeline from unposed day-to-day images is equally crucial in neural novel view synthesis problem; (ii) It is rather more practical to model object’s content at different resolutions since dramatic camera motion is highly likely in day-to-day unposed images. To incorporate the key ideas, we leverage the fundamentals of scene rigidity, multi-scale neural scene representation, and single-image depth prediction. Concretely, the proposed approach makes the camera parameters as learnable in a neural fields-based modeling framework. By assuming per view depth prediction is given up to scale, we constrain the relative pose between successive frames. From the relative poses, absolute camera pose estimation is modeled via a graph-neural network-based multiple motion averaging within the multi-scale neural-fields network, leading to a single loss function. Optimizing the introduced loss function provides camera intrinsic, extrinsic, and image rendering from unposed images. We demonstrate, with examples, that for a unified framework to accurately model multiscale neural scene representation from day-to-day acquired unposed multi-view images, it is equally essential to have precise camera-pose estimates within the scene representation framework. Without considering robustness measures in the camera pose estimation pipeline, modeling for multi-scale aliasing artifacts can be counterproductive. We present extensive experiments on several benchmark datasets to demonstrate the suitability of our approach.

Federated learning (FL) is a hot collaborative training framework via aggregating model parameters of decentralized local clients. However, most FL methods unreasonably assume data categories of FL framework are known and fixed in advance. Moreover, some new local clients that collect novel categories unseen by other clients may be introduced to FL training irregularly. These issues render global model to undergo catastrophic forgetting on old categories, when local clients receive new categories consecutively under limited memory of storing old categories. To tackle the above issues, we propose a novelLocal-GlobalAnti-forgetting (LGA) model. It ensures no local clients are left behind as they learn new classes continually, by addressing local and global catastrophic forgetting. Specifically, considering tackling class imbalance of local client to surmount local forgetting, we develop a category-balanced gradient-adaptive compensation loss and a category gradient-induced semantic distillation loss. They can balance heterogeneous forgetting speeds of hard-to-forget and easy-to-forget old categories, while ensure consistent class-relations within different tasks. Moreover, a proxy server is designed to tackle global forgetting caused by Non-IID class imbalance between different clients. It augments perturbed prototype images of new categories collected from local clients via self-supervised prototype augmentation, thus improving robustness to choose the best old global model for local-side semantic distillation loss. Experiments on representative datasets verify superior performance of our model against comparison methods. The code is available athttps://github.com/JiahuaDong/LGA.

Unsupervised domain adaptation (UDA) and domain generalization (DG) enable machine learning models trained on a source domain to perform well on unlabeled or even unseen target domains. As previous UDA&DG semantic segmentation methods are mostly based on outdated networks, we benchmark more recent architectures, reveal the potential of Transformers, and design the DAFormer network tailored for UDA&DG. It is enabled by three training strategies to avoid overfitting to the source domain: While (1) Rare Class Sampling mitigates the bias toward common source domain classes, (2) a Thing-Class ImageNet Feature Distance and (3) a learning rate warmup promote feature transfer from ImageNet pretraining. As UDA&DG are usually GPU memory intensive, most previous methods downscale or crop images. However, low-resolution predictions often fail to preserve fine details while models trained with cropped images fall short in capturing long-range, domain-robust context information. Therefore, we propose HRDA, a multi-resolution framework for UDA&DG, that combines the strengths of small high-resolution crops to preserve fine segmentation details and large low-resolution crops to capture long-range context dependencies with a learned scale attention. DAFormer and HRDA significantly improve the state-of-the-art UDA&DG by more than 10 mIoU on 5 different benchmarks.

Contextual information plays a core role for video semantic segmentation (VSS). This paper summarizes contexts for VSS in two-fold: local temporal contexts (LTC) which define the contexts from neighboring frames, and global temporal contexts (GTC) which represent the contexts from the whole video. As for LTC, it includes static and motional contexts, corresponding to static and moving content in neighboring frames, respectively. Previously, both static and motional contexts have been studied. However, there is no research about simultaneously learning static and motional contexts (highly complementary). Hence, we propose a Coarse-to-Fine Feature Mining (CFFM) technique to learn a unified presentation of LTC. CFFM contains two parts: Coarse-to-Fine Feature Assembling (CFFA) and Cross-frame Feature Mining (CFM). CFFA abstracts static and motional contexts, and CFM mines useful information from nearby frames to enhance target features. To further exploit more temporal contexts, we propose CFFM++ by additionally learning GTC from the whole video. Specifically, we uniformly sample certain frames from the video and extract global contextual prototypes by k-means. The information within those prototypes is mined by CFM to refine target features. Experimental results on popular benchmarks demonstrate that CFFM and CFFM++ perform favorably against state-of-the-art methods.

We present a novel graph Transformer generative adversarial network (GTGAN) to learn effective graph node relations in an end-to-end fashion for challenging graph-constrained architectural layout generation tasks. The proposed graph-Transformer-based generator includes a novel graph Transformer encoder that combines graph convolutions and self-attentions in a Transformer to model both local and global interactions across connected and non-connected graph nodes. Specifically, the proposed connected node attention (CNA) and non-connected node attention (NNA) aim to capture the global relations across connected nodes and non-connected nodes in the input graph, respectively. The proposed graph modeling block (GMB) aims to exploit local vertex interactions based on a house layout topology. Moreover, we propose a new node classification-based discriminator to preserve the high-level semantic and discriminative node features for different house components. To maintain the relative spatial relationships between ground truth and predicted graphs, we also propose a novel graph-based cycle-consistency loss. Finally, we propose a novel self-guided pre-training method for graph representation learning. This approach involves simultaneous masking of nodes and edges at an elevated mask ratio (i.e., 40%) and their subsequent reconstruction using an asymmetric graph-centric autoencoder architecture. This method markedly improves the model's learning proficiency and expediency. Experiments on three challenging graph-constrained architectural layout generation tasks (i.e., house layout generation, house roof generation, and building layout generation) with three public datasets demonstrate the effectiveness of the proposed method in terms of objective quantitative scores and subjective visual realism. New state-of-the-art results are established by large margins on these three tasks.

This article studies the problem of learning weakly supervised semantic segmentation (WSSS) from image-level supervision only. Current popular solutions leverage object localization maps from classifiers as supervision for semantic segmentation learning, and struggle to make the localization maps capture more complete object content. Rather than previous efforts that primarily focus on intra-image information, we address the value of cross-image semantic relations for comprehensive object pattern mining. To achieve this, two neural co-attentions are incorporated into the classifier to complementarily capture cross-image semantic similarities and differences. In particular, given a pair of training images, one co-attention enforces the classifier to recognize the common semantics from co-attentive objects, while the other one, called contrastive co-attention, drives the classifier to identify the unique semantics from the rest, unshared objects. This helps the classifier discover more object patterns and better ground semantics in image regions. In addition to boosting object pattern learning, the co-attention can leverage context from other related images to improve localization map inference, hence eventually benefiting semantic segmentation learning. More importantly, our algorithm provides a unified framework that handles well different WSSS settings, i.e., learning WSSS with 1) precise image-level supervision only, 2) extra simple single-label data, and 3) extra noisy web data. Without bells and whistles, it sets new state-of-the-arts on all these settings. Moreover, our approach ranked 1st place in the Weakly-Supervised Semantic Segmentation Track of CVPR2020 Learning from Imperfect Data Challenge. The extensive experimental results demonstrate well the efficacy and high utility of our method.

Despite the impressive achievements of Deep Neural Networks (DNNs) in computer vision, their vulnerability to adversarial attacks remains a critical concern. Extensive research has demonstrated that incorporating sophisticated perturbations into input images can lead to a catastrophic degradation in DNNs' performance. This perplexing phenomenon not only exists in the digital space but also in the physical world. Consequently, it becomes imperative to evaluate the security of DNNs-based systems to ensure their safe deployment in real-world scenarios, particularly in security-sensitive applications. To facilitate a profound understanding of this topic, this paper presents a comprehensive overview of physical adversarial attacks. First, we distill four general steps for launching physical adversarial attacks. Building upon this foundation, we uncover the pervasive role of artifacts carrying adversarial perturbations in the physical world. These artifacts influence each step. To denote them, we introduce a new term: adversarial medium. Then, we take the first step to systematically evaluate the performance of physical adversarial attacks, taking the adversarial medium as a first attempt. Our proposed evaluation metric, hiPAA, comprises six perspectives: Effectiveness, Stealthiness, Robustness, Practicability, Aesthetics, and Economics. We also provide comparative results across task categories, together with insightful observations and suggestions for future research directions.

How to identify and segment camouflaged objects from the background is challenging. Inspired by the multi-head self-attention in Transformers, we present a simple masked separable attention (MSA) for camouflaged object detection. We first separate the multi-head self-attention into three parts, which are responsible for distinguishing the camouflaged objects from the background using different mask strategies. Furthermore, we propose to capture high-resolution semantic representations progressively based on a simple top-down decoder with the proposed MSA to attain precise segmentation results. These structures plus a backbone encoder form a new model, dubbed CamoFormer. Extensive experiments show that CamoFormer achieves new state-of-the-art performance on three widely-used camouflaged object detection benchmarks. To better evaluate the performance of the proposed CamoFormer around the border regions, we propose to use two new metrics, i.e., BR-M and BR-F. There are on average ∼ 5% relative improvements over previous methods in terms of S-measure and weighted F-measure.

High Dynamic Range (HDR) images can be recovered from several Low Dynamic Range (LDR) images by existing Deep Neural Networks (DNNs) techniques. Despite the remarkable progress, DNN-based methods still generate ghosting artifacts when LDR images have saturation and large motion, which hinders potential applications in real-world scenarios. To address this challenge, we formulate the HDR deghosting problem as an image generation that leverages LDR features as the diffusion model’s condition, consisting of the feature condition generator and the noise predictor. Feature condition generator employs attention and Domain Feature Alignment (DFA) layer to transform the intermediate features to avoid ghosting artifacts. With the learned features as conditions, the noise predictor leverages a stochastic iterative denoising process for diffusion models to generate an HDR image by steering the sampling process. Furthermore, to mitigate semantic confusion caused by the saturation problem of LDR images, we design a sliding window noise estimator to sample smooth noise in a patch-based manner. In addition, an image space loss is proposed to avoid the color distortion of the estimated HDR results. We empirically evaluate our model on benchmark datasets for HDR imaging. The results demonstrate that our approach achieves state-of-the-art performances and well generalization to real-world images.

Video restoration aims to restore high-quality frames from low-quality frames. Different from single image restoration, video restoration generally requires to utilize temporal information from multiple adjacent but usually misaligned video frames. Existing deep methods generally tackle with this by exploiting a sliding window strategy or a recurrent architecture, which are restricted by frame-by-frame restoration. In this paper, we propose a Video Restoration Transformer (VRT) with parallel frame prediction ability. More specifically, VRT is composed of multiple scales, each of which consists of two kinds of modules: temporal reciprocal self attention (TRSA) and parallel warping. TRSA divides the video into small clips, on which reciprocal attention is applied for joint motion estimation, feature alignment and feature fusion, while self attention is used for feature extraction. To enable cross-clip interactions, the video sequence is shifted for every other layer. Besides, parallel warping is used to further fuse information from neighboring frames by parallel feature warping. Experimental results on five tasks, including video super-resolution, video deblurring, video denoising, video frame interpolation and space-time video super-resolution, demonstrate that VRT outperforms the state-of-the-art methods by large margins (up to 2.16dB) on fourteen benchmark datasets. The codes are available at https://github.com/JingyunLiang/VRT.

We present a novel graph Transformer generative adversarial network (GTGAN) to learn effective graph node relations in an end-to-end fashion for the challenging graph-constrained house generation task. The proposed graph-Transformer-based generator includes a novel graph Transformer encoder that combines graph convolutions and self-attentions in a Transformer to model both local and global interactions across connected and non-connected graph nodes. Specifically, the proposed connected node attention (CNA) and non-connected node attention (NNA) aim to capture the global relations across connected nodes and non-connected nodes in the input graph, respectively. The proposed graph modeling block (GMB) aims to exploit local vertex interactions based on a house layout topology. More-over, we propose a new node classification-based discriminator to preserve the high-level semantic and discriminative node features for different house components. Finally, we propose a novel graph-based cycle-consistency loss that aims at maintaining the relative spatial relationships between ground truth and predicted graphs. Experiments on two challenging graph-constrained house generation tasks (i.e., house layout and roof generation) with two public datasets demonstrate the effectiveness of GTGAN in terms of objective quantitative scores and subjective visual realism. New state-of-the-art results are established by large margins on both tasks.

Learning continuous image representations is recently gaining popularity for image super-resolution (SR) because of its ability to reconstruct high-resolution images with arbitrary scales from low-resolution inputs. Existing methods mostly ensemble nearby features to predict the new pixel at any queried coordinate in the SR image. Such a local ensemble suffers from some limitations: i) it has no learnable parameters and it neglects the similarity of the visual features; ii) it has a limited receptive field and cannot ensemble relevant features in a large field which are important in an image. To address these issues, this paper proposes a continuous implicit attention-in-attention network, called CiaoSR. We explicitly design an implicit attention network to learn the ensemble weights for the nearby local features. Furthermore, we embed a scale-aware attention in this implicit attention network to exploit additional non-local information. Extensive experiments on benchmark datasets demonstrate CiaoSR significantly outperforms the existing single image SR methods with the same backbone. In addition, CiaoSR also achieves the state-of-the-art performance on the arbitrary-scale SR task. The effectiveness of the method is also demonstrated on the real-world SR setting. More importantly, CiaoSR can be flexibly integrated into any backbone to improve the SR performance.

Unsupervised domain adaptation (UDA) for semantic segmentation aims at improving the model performance on the unlabeled target domain by leveraging a labeled source domain. Existing approaches have achieved impressive progress by utilizing pseudo-labels on the unlabeled target-domain images. Yet the low-quality pseudo-labels, arising from the domain discrepancy, inevitably hinder the adaptation. This calls for effective and accurate approaches to estimating the reliability of the pseudo-labels, in order to rectify them. In this paper, we propose to estimate the rectification values of the predicted pseudo-labels with implicit neural representations. We view the rectification value as a signal defined over the continuous spatial domain. Taking an image coordinate and the nearby deep features as inputs, the rectification value at a given coordinate is predicted as an output. This allows us to achieve high-resolution and detailed rectification values estimation, important for accurate pseudo-label generation at mask boundaries in particular. The rectified pseudo-labels are then leveraged in our rectification-aware mixture model (RMM) to be learned end-to-end and help the adaptation. We demonstrate the effectiveness of our approach on different UDA benchmarks, including synthetic-to-real and day-to-night. Our approach achieves superior results compared to state-of-the-art. The implementation is available at https://github.com/ETHRuiGong/IR2F.

In unsupervised domain adaptation (UDA), a model trained on source data (e.g. synthetic) is adapted to target data (e.g. real-world) without access to target annotation. Most previous UDA methods struggle with classes that have a similar visual appearance on the target domain as no ground truth is available to learn the slight appearance differences. To address this problem, we propose a Masked Image Consistency (MIC) module to enhance UDA by learning spatial context relations of the target domain as additional clues for robust visual recognition. MIC enforces the consistency between predictions of masked target images, where random patches are withheld, and pseudo-labels that are generated based on the complete image by an exponential moving average teacher. To minimize the consistency loss, the network has to learn to infer the predictions of the masked regions from their context. Due to its simple and universal concept, MIC can be integrated into various UDA methods across different visual recognition tasks such as image classification, semantic segmentation, and object detection. MIC significantly improves the state-of-the-art performance across the different recognition tasks for synthetic-to-real, day-to-nighttime, and clear-toadverse-weather UDA. For instance, MIC achieves an un-precedented UDA performance of 75.9 mIoU and 92.8% on GTA→Cityscapes and VisDA-2017, respectively, which corresponds to an improvement of +2.1 and +3.0 percent points over the previous state of the art. The implementation is available at https://github.com/lhoyer/MIC.

We introduce an approach to enhance the novel view synthesis from images taken from a freely moving camera. The introduced approach focuses on outdoor scenes where recovering accurate geometric scaffold and camera pose is challenging, leading to inferior results using the state-of-the-art stable view synthesis (SVS) method. SVS and related methods fail for outdoor scenes primarily due to (i) overrelying on the multiview stereo (MVS) for geometric scaffold recovery and (ii) assuming COLMAP computed camera poses as the best possible estimates, despite it being well-studied that MVS 3D reconstruction accuracy is limited to scene disparity and camera-pose accuracy is sensitive to key-point correspondence selection. This work proposes a principled way to enhance novel view synthesis solutions drawing inspiration from the basics of multiple view geometry. By leveraging the complementary behavior of MVS and monocular depth, we arrive at a better scene depth per view for nearby and far points, respectively. Moreover, our approach jointly refines camera poses with image-based rendering via multiple rotation averaging graph optimization. The recovered scene depth and the camera-pose help better view-dependent on-surface feature aggregation of the entire scene. Extensive evaluation of our approach on the popular benchmark dataset, such as Tanks and Temples, shows substantial improvement in view synthesis results compared to the prior art. For instance, our method shows 1.5 dB of PSNR improvement on the Tank and Temples. Similar statistics are observed when tested on other benchmark datasets such as FVS, Mip-NeRF 360, and DTU.

The aim of this paper is to propose a mechanism to efficiently and explicitly model image hierarchies in the global, regional, and local range for image restoration. To achieve that, we start by analyzing two important properties of natural images including cross-scale similarity and anisotropic image features. Inspired by that, we propose the anchored stripe self-attention which achieves a good balance between the space and time complexity of self-attention and the modelling capacity beyond the regional range. Then we propose a new network architecture dubbed GRL to explicitly model image hierarchies in the Global, Regional, and Local range via anchored stripe self-attention, window self-attention, and channel attention enhanced convolution. Finally, the proposed network is applied to 7 image restoration types, covering both real and synthetic settings. The proposed method sets the new state-of-the-art for several of those. Code will be available at https://github.com/ofsoundof/GRL-Image-Restoration.git.

Neural-network-based single image depth prediction (SIDP) is a challenging task where the goal is to predict the scene's per-pixel depth at test time. Since the problem, by definition, is illposed, the fundamental goal is to come up with an approach that can reliably model the scene depth from a set of training examples. In the pursuit of perfect depth estimation, most existing state-of-the-art learning techniques predict a single scalar depth value per-pixel. Yet, it is well-known that the trained model has accuracy limits and can predict imprecise depth. Therefore, an SIDP approach must be mindful of the expected depth variations in the model's prediction at test time. Accordingly, we introduce an approach that performs continuous modeling of per-pixel depth, where we can predict and reason about the per-pixel depth and its distribution. To this end, we model per-pixel scene depth using a multivariate Gaussian distribution. Moreover, contrary to the existing uncertainty modeling methods—in the same spirit, where per-pixel depth is assumed to be independent, we introduce per-pixel covariance modeling that encodes its depth dependency w.r.t. all the scene points. Unfortunately, per-pixel depth covariance modeling leads to a computationally expensive continuous loss function, which we solve efficiently using the learned low-rank approximation of the overall covariance matrix. Notably, when tested on benchmark datasets such as KITTI, NYU, and SUN-RGB-D, the SIDP model obtained by optimizing our loss function shows state-of-the-art results. Our method's accuracy (named MG) is among the top on the KITTI depth-prediction benchmark leaderboard11http://www.cvlibs.net/datasets/kitti/eval–depth.php?benchmark=depth–prediction.

Deep point cloud registration methods face challenges to partial overlaps and rely on labeled data. To address these issues, we propose UDPReg, an unsupervised deep probabilistic registration framework for point clouds with partial overlaps. Specifically, we first adopt a network to learn posterior probability distributions of Gaussian mixture models (GMMs) from point clouds. To handle partial point cloud registration, we apply the Sinkhorn algorithm to predict the distribution-level correspondences under the constraint of the mixing weights of GMMs. To enable unsupervised learning, we design three distribution consistency-based losses: self-consistency, cross-consistency, and local contrastive. The self-consistency loss is formulated by encouraging GMMs in Euclidean and feature spaces to share identical posterior distributions. The cross-consistency loss derives from the fact that the points of two partially overlapping point clouds belonging to the same clusters share the cluster centroids. The cross-consistency loss allows the network to flexibly learn a transformation-invariant posterior distribution of two aligned point clouds. The local contrastive loss facilitates the network to extract discriminative local features. Our UDPReg achieves competitive performance on the 3DMatch/3DLoMatch and ModelNet/ModelLoNet benchmarks.

Recent works have shown that unstructured text (doc-uments) from online sources can serve as useful auxiliary information for zero-shot image classification. However, these methods require access to a high-quality source like Wikipedia and are limited to a single source of information. Large Language Models (LLM) trained on web-scale text show impressive abilities to repurpose their learned knowledge for a multitude of tasks. In this work, we provide a novel perspective on using an LLM to provide text supervision for a zero-shot image classification model. The LLM is provided with a few text descriptions from different annota-tors as examples. The LLM is conditioned on these exam-ples to generate multiple text descriptions for each class (re-ferred to as views). Our proposed model, I2MVFormer, learns multi-view semantic embeddings for zero-shot image classification with these class views. We show that each text view of a class provides complementary information allowing a model to learn a highly discriminative class embed-ding. Moreover, we show that I2MVFormer is better at consuming the multi-view text supervision from LLM compared to baseline models. I2MVFormer establishes a new state-of-the-art on three public benchmark datasets for zero-shot image classification with unsupervised semantic embeddings. Code available at https://github.com/ferjad/I2DFormer

During training, supervised object detection tries to correctly match the predicted bounding boxes and associated classification scores to the ground truth. This is essential to determine which predictions are to be pushed towards which solutions, or to be discarded. Popular matching strategies include matching to the closest ground truth box (mostly used in combination with anchors), or matching via the Hungarian algorithm (mostly used in anchor free methods). Each of these strategies comes with its own properties, underlying losses, and heuristics. We show how Unbalanced Optimal Transport unifies these different approaches and opens a whole continuum of methods in between. This allows for a finer selection of the desired properties. Experimentally, we show that training an object detection model with Unbalanced Optimal Transport is able to reach the state-of-the-art both in terms of Average Precision and Average Recall as well as to provide a faster initial convergence. The approach is well suited for GPU implementation, which proves to be an advantage for large-scale models.

Recently, indiscernible scene understanding has attracted a lot of attention in the vision community. We further advance the frontier of this field by systematically studying a new challenge named indiscernible object counting (IOC), the goal of which is to count objects that are blended with respect to their surroundings. Due to a lack of appropriate IOC datasets, we present a large-scale dataset IOCfish5K which contains a total of 5,637 high-resolution images and 659,024 annotated center points. Our dataset consists of a large number of indiscernible objects (mainly fish) in underwater scenes, making the annotation process all the more challenging. IOCfish5K is superior to existing datasets with indiscernible scenes because of its larger scale, higher image resolutions, more annotations, and denser scenes. All these aspects make it the most challenging dataset for IOC so far, supporting progress in this area. For benchmarking purposes, we select 14 mainstream methods for object counting and carefully evaluate them on IOCfish5K. Furthermore, we propose IOCFormer, a new strong baseline that combines density and regression branches in a unified framework and can effectively tackle object counting under concealed scenes. Experiments show that IOCFormer achieves state-of-the-art scores on IOCfish5K. The resources are available at github.com/GuoleiSun/Indiscernible-Object-Counting.

The performance of video frame interpolation is inherently correlated with the ability to handle motion in the input scene. Even though previous works recognize the utility of asynchronous event information for this task, they ignore the fact that motion may or may not result in blur in the input video to be interpolated, depending on the length of the exposure time of the frames and the speed of the motion, and assume either that the input video is sharp, restricting themselves to frame interpolation, or that it is blurry, including an explicit, separate deblurring stage before interpolation in their pipeline. We instead propose a general method for event-based frame interpolation that performs deblurring ad-hoc and thus works both on sharp and blurry input videos. Our model consists in a bidirectional recurrent network that naturally incorporates the temporal dimension of interpolation and fuses information from the input frames and the events adaptively based on their temporal proximity. In addition, we introduce a novel real-world high-resolution dataset with events and color videos named HighREV, which provides a challenging evaluation setting for the examined task. Extensive experiments on the standard CoPro benchmark and on our dataset show that our network consistently outperforms previous state-of-the-art methods on frame interpolation, single image deblurring and the joint task of interpolation and deblurring. Our code and dataset are available at https://github.com/AHupuJR/REFID.

The high computational costs of video super-resolution (VSR) models hinder their deployment on resource-limited devices, e.g., smartphones and drones. Existing VSR models contain considerable redundant filters, which drag down the inference efficiency. To prune these unimportant filters, we develop a structured pruning scheme called Structured Sparsity Learning (SSL) according to the properties of VSR. In SSL, we design pruning schemes for several key components in VSR models, including residual blocks, recurrent networks, and upsampling networks. Specifically, we develop a Residual Sparsity Connection (RSC) scheme for residual blocks of recurrent networks to liberate pruning restrictions and preserve the restoration information. For upsampling networks, we design a pixel-shuffle pruning scheme to guarantee the accuracy of feature channel-space conversion. In addition, we observe that pruning error would be amplified as the hidden states propagate along with recurrent networks. To alleviate the issue, we design Temporal Finetuning (TF). Extensive experiments show that SSL can significantly outperform recent methods quantitatively and qualitatively. The code is available at https://github.com/Zj-BinXia/SSL.

Generating a high-quality High Dynamic Range (HDR) image from dynamic scenes has recently been extensively studied by exploiting Deep Neural Networks (DNNs). Most DNNs-based methods require a large amount of training data with ground truth, requiring tedious and time-consuming work. Few-shot HDR imaging aims to generate satisfactory images with limited data. However, it is difficult for modern DNNs to avoid overfitting when trained on only a few images. In this work, we propose a novel semi-supervised approach to realize few-shot HDR imaging via two stages of training, called SSHDR. Unlikely previous methods, directly recovering content and removing ghosts simultaneously, which is hard to achieve optimum, we first generate content of saturated regions with a self-supervised mechanism and then address ghosts via an iterative semi-supervised learning framework. Concretely, considering that saturated regions can be regarded as masking Low Dynamic Range (LDR) input regions, we design a Saturated Mask AutoEncoder (SMAE) to learn a robust feature representation and reconstruct a non-saturated HDR image. We also propose an adaptive pseudo-label selection strategy to pick high-quality HDR pseudo-labels in the second stage to avoid the effect of mislabeled samples. Experiments demonstrate that SSHDR outperforms state-of-the-art methods quantitatively and qualitatively within and across different datasets, achieving appealing HDR visualization with few labeled samples.

Multi-modality (MM) image fusion aims to render fused images that maintain the merits of different modalities, e.g., functional highlight and detailed textures. To tackle the challenge in modeling cross-modality features and decomposing desirable modality-specific and modality-shared features, we propose a novel Correlation-Driven feature Decomposition Fusion (CDDFuse) network. Firstly, CDDFuse uses Restormer blocks to extract cross-modality shallow features. We then introduce a dual-branch Transformer-CNN feature extractor with Lite Transformer (LT) blocks leveraging long-range attention to handle low-frequency global features and Invertible Neural Networks (INN) blocks focusing on extracting high-frequency local information. A correlation-driven loss is further proposed to make the low-frequency features correlated while the high-frequency features uncorrelated based on the embedded information. Then, the LT-based global fusion and INN-based local fusion layers output the fused image. Extensive experiments demonstrate that our CDDFuse achieves promising results in multiple fusion tasks, including infrared-visible image fusion and medical image fusion. We also show that CDDFuse can boost the performance in downstream infrared-visible semantic segmentation and object detection in a unified benchmark. The code is available at https://github.om/haozixiang1228/MMIF-CDDFuse.

With autonomous industries on the rise, domain adaptation of the visual perception stack is an important research direction due to the cost savings promise. Much prior art was dedicated to domain-adaptive semantic segmentation in the synthetic-to-real context. Despite being a crucial output of the perception stack, panoptic segmentation has been largely overlooked by the domain adaptation community. Therefore, we revisit well-performing domain adaptation strategies from other fields, adapt them to panoptic segmentation, and show that they can effectively enhance panoptic domain adaptation. Further, we study the panoptic network design and propose a novel architecture (EDAPS) designed explicitly for domain-adaptive panoptic segmentation. It uses a shared, domain-robust transformer encoder to facilitate the joint adaptation of semantic and instance features, but task-specific decoders tailored for the specific requirements of both domain-adaptive semantic and instance segmentation. As a result, the performance gap seen in challenging panoptic benchmarks is substantially narrowed. EDAPS significantly improves the state-of-the-art performance for panoptic segmentation UDA by a large margin of 20% on SYNTHIA-to-Cityscapes and even 72% on the more challenging SYNTHIA-to-Mapillary Vistas. The implementation is available at https://github.com/susaha/edaps.

We introduce a self-supervised training strategy for burst super-resolution that only uses noisy low-resolution bursts during training. Our approach eliminates the need to carefully tune synthetic data simulation pipelines, which often do not match real-world image statistics. Compared to weakly-paired training strategies, which require noisy smartphone burst photos of static scenes, paired with a clean reference obtained from a tripod-mounted DSLR camera, our approach is more scalable, and avoids the color mismatch between the smartphone and DSLR. To achieve this, we propose a new self-supervised objective that uses a forward imaging model to recover a high-resolution image from aliased high frequencies in the burst. Our approach does not require any manual tuning of the forward model's parameters; we learn them from data. Furthermore, we show our training strategy is robust to dynamic scene motion in the burst, which enables training burst super-resolution models using in-the-wild data. Extensive experiments on real and synthetic data show that, despite only using noisy bursts during training, models trained with our self-supervised strategy match, and sometimes surpass, the quality of fully-supervised baselines trained with synthetic data or weakly-paired ground-truth. Finally, we show our training strategy is general using four different burst super-resolution architectures.

Standard unsupervised domain adaptation methods adapt models from a source to a target domain using labeled source data and unlabeled target data jointly. In model adaptation, on the other hand, access to the labeled source data is prohibited, i.e., only the source-trained model and unlabeled target data are available. We investigate normal-to-adverse condition model adaptation for semantic segmentation, whereby image-level correspondences are available in the target domain. The target set consists of unlabeled pairs of adverse- and normal-condition street images taken at GNSS-matched locations. Our method—CMA—leverages such image pairs to learn condition-invariant features via contrastive learning. In particular, CMA encourages features in the embedding space to be grouped according to their condition-invariant semantic content and not according to the condition under which respective inputs are captured. To obtain accurate cross-domain semantic correspondences, we warp the normal image to the viewpoint of the adverse image and leverage warp-confidence scores to create robust, aggregated features. With this approach, we achieve state-of-the-art semantic segmentation performance for model adaptation on several normal-to-adverse adaptation benchmarks, such as ACDC and Dark Zurich. We also evaluate CMA on a newly procured adverse-condition generalization benchmark and report favorable results compared to standard unsupervised domain adaptation methods, despite the comparative handicap of CMA due to source data inaccessibility. Code is available at https://github.com/brdav/cma.

Scene extrapolation—the idea of generating novel views by flying into a given image—is a promising, yet challenging task. For each predicted frame, a joint inpainting and 3D refinement problem has to be solved, which is ill posed and includes a high level of ambiguity. Moreover, training data for long-range scenes is difficult to obtain and usually lacks sufficient views to infer accurate camera poses. We introduce DiffDreamer, an unsupervised framework capable of synthesizing novel views depicting a long camera trajectory while training solely on internet-collected images of nature scenes. Utilizing the stochastic nature of the guided denoising steps, we train the diffusion models to refine projected RGBD images but condition the denoising steps on multiple past and future frames for inference. We demonstrate that image-conditioned diffusion models can effectively perform long-range scene extrapolation while preserving consistency significantly better than prior GAN-based methods. DiffDreamer is a powerful and efficient solution for scene extrapolation, producing impressive results despite limited supervision. Project page: https://primecai.github.io/diffdreamer.

Autonomous driving requires accurate local scene understanding information. To this end, autonomous agents deploy object detection and online BEV lane graph extraction methods as a part of their perception stack. In this work, we propose an architecture and loss formulation to improve the accuracy of local lane graph estimates by using 3D object detection outputs. The proposed method learns to assign the objects to centerlines by considering the centerlines as cluster centers and the objects as data points to be assigned a probability distribution over the cluster centers. This training scheme ensures direct supervision on the relationship between lanes and objects, thus leading to better performance. The proposed method improves lane graph estimation substantially over state-of-the-art methods. The extensive ablations show that our method can achieve significant performance improvements by using the outputs of existing 3D object detection methods. Since our method uses the detection outputs rather than detection method intermediate representations, a single model of our method can use any detection method at test time. The code will be made publicly available.

Performing multiple heterogeneous visual tasks in dynamic scenes is a hallmark of human perception capability. Despite remarkable progress in image and video recognition via representation learning, current research still focuses on designing specialized networks for singular, homogeneous, or simple combination of tasks. We instead explore the construction of a unified model for major image and video recognition tasks in autonomous driving with diverse input and output structures. To enable such an investigation, we design a new challenge, Video Task Decathlon (VTD), which includes ten representative image and video tasks spanning classification, segmentation, localization, and association of objects and pixels. On VTD, we develop our unified network, VTDNet, that uses a single structure and a single set of weights for all ten tasks. VTDNet groups similar tasks and employs task interaction stages to exchange information within and between task groups. Given the impracticality of labeling all tasks on all frames and the performance degradation associated with joint training of many tasks, we design a Curriculum training, Pseudo-labeling, and Fine-tuning (CPF) scheme to successfully train VTDNet on all tasks and mitigate performance loss. Armed with CPF, VTDNet significantly outperforms its single-task counterparts on most tasks with only 20% overall computations. VTD is a promising new direction for exploring the unification of perception tasks in autonomous driving.

Continual Learning aims to learn a single model on a sequence of tasks without having access to data from previous tasks. The biggest challenge in the domain still remains catastrophic forgetting: a loss in performance on seen classes of earlier tasks. Some existing methods rely on an expensive replay buffer to store a chunk of data from previous tasks. This, while promising, becomes expensive when the number of tasks becomes large or data can not be stored for privacy reasons. As an alternative, prompt-based methods have been proposed that store the task information in a learnable prompt pool. This prompt pool instructs a frozen image encoder on how to solve each task. While the model faces a disjoint set of classes in each task in this setting, we argue that these classes can be encoded to the same embedding space of a pre-trained language encoder. In this work, we propose Language Guidance for Prompt-based Continual Learning (LGCL) as a plug-in for prompt-based methods. LGCL is model agnostic and introduces language guidance at the task level in the prompt pool and at the class level on the output feature of the vision encoder. We show with extensive experimentation that LGCL consistently improves the performance of prompt-based continual learning methods to set a new state-of-the art. LGCL achieves these performance improvements without needing any additional learnable parameters.

Modeling Neural Radiance Fields for fast-moving deformable objects from visual data alone is a challenging problem. A major issue arises due to the high deformation and low acquisition rates. To address this problem, we propose to use event cameras that offer very fast acquisition of visual change in an asynchronous manner. In this work, we develop a novel method to model the deformable neural radiance fields using RGB and event cameras. The proposed method uses the asynchronous stream of events and calibrated sparse RGB frames. In our setup, the camera pose at the individual events –required to integrate them into the radiance fields– remains unknown. Our method jointly optimizes these poses and the radiance field. This happens efficiently by leveraging the collection of events at once and actively sampling the events during learning. Experiments conducted on both realistically rendered graphics and real-world datasets demonstrate a significant benefit of the proposed method over the state-of-the-art and the compared baseline. This shows a promising direction for modeling deformable neural radiance fields in real-world dynamic scenes. We release our code at: https://qimaqi.github.io/DE-NeRF.github.io/

Novel view synthesis and 3D modeling using implicit neural field representation are shown to be very effective for calibrated multi-view cameras. Such representations are known to benefit from additional geometric and semantic supervision. Most existing methods that exploit additional supervision require dense pixel-wise labels or localized scene priors. These methods cannot benefit from high-level vague scene priors provided in terms of scenes’ descriptions. In this work, we aim to leverage the geometric prior of Manhattan scenes to improve the implicit neural radiance field representations. More precisely, we assume that only the knowledge of the indoor scene (under investigation) being Manhattan is known – with no additional information whatsoever – with an unknown Manhattan coordinate frame. Such high-level prior is used to self-supervise the surface normals derived explicitly in the implicit neural fields. Our modeling allows us to cluster the derived normals and exploit their orthogonality constraints for self-supervision. Our exhaustive experiments on datasets of diverse indoor scenes demonstrate the significant benefit of the proposed method over the established baselines. The source code will be available at https://github.com/nikola3794/normal-clustering-nerf.

We propose a dense neural simultaneous localization and mapping (SLAM) approach for monocular RGBD input which anchors the features of a neural scene representation in a point cloud that is iteratively generated in an input-dependent data-driven manner. We demonstrate that both tracking and mapping can be performed with the same point-based neural scene representation by minimizing an RGBD-based re-rendering loss. In contrast to recent dense neural SLAM methods which anchor the scene features in a sparse grid, our point-based approach allows dynamically adapting the anchor point density to the information density of the input. This strategy reduces runtime and memory usage in regions with fewer details and dedicates higher point density to resolve fine details. Our approach performs either better or competitive to existing dense neural RGBD SLAM methods in tracking, mapping and rendering accuracy on the Replica, TUM-RGBD and Scan-Net datasets. The source code is available at https://github.com/eriksandstroem/Point-SLAM.

VLN-CE is a recently released embodied task, where AI agents need to navigate a freely traversable environment to reach a distant target location, given language instructions. It poses great challenges due to the huge space of possible strategies. Driven by the belief that the ability to anticipate the consequences of future actions is crucial for the emergence of intelligent and interpretable planning behavior, we propose Dreamwalker — a world model based VLN-CE agent. The world model is built to summarize the visual, topological, and dynamic properties of the complicated continuous environment into a discrete, structured, and compact representation. Dreamwalker can simulate and evaluate possible plans entirely in such internal abstract world, before executing costly actions. As opposed to existing model-free VLN-CE agents simply making greedy decisions in the real world, which easily results in shortsighted behaviors, Dreamwalker is able to make strategic planning through large amounts of "mental experiments." Moreover, the imagined future scenarios reflect our agent’s intention, making its decision-making process more transparent. Extensive experiments and ablation studies on VLN-CE dataset confirm the effectiveness of the proposed approach and outline fruitful directions for future work.

Depth cues are known to be useful for visual perception. However, direct measurement of depth is often impracticable. Fortunately, though, modern learning-based methods offer promising depth maps by inference in the wild. In this work, we adapt such depth inference models for object segmentation using the objects’ "pop-out" prior in 3D. The "pop-out" is a simple composition prior that assumes objects reside on the background surface. Such compositional prior allows us to reason about objects in the 3D space. More specifically, we adapt the inferred depth maps such that objects can be localized using only 3D information. Such separation, however, requires knowledge about contact surface which we learn using the weak supervision of the segmentation mask. Our intermediate representation of contact surface, and thereby reasoning about objects purely in 3D, allows us to better transfer the depth knowledge into semantics. The proposed adaptation method uses only the depth model without needing the source data used for training, making the learning process efficient and practical. Our experiments on eight datasets of two challenging tasks, namely salient object detection and camouflaged object detection, consistently demonstrate the benefit of our method in terms of both performance and generalizability. The source code is publicly available at https://github.com/Zongwei97/PopNet.

Diffusion model (DM) has achieved SOTA performance by modeling the image synthesis process into a sequential application of a denoising network. However, different from image synthesis, image restoration (IR) has a strong constraint to generate results in accordance with ground-truth. Thus, for IR, traditional DMs running massive iterations on a large model to estimate whole images or feature maps is inefficient. To address this issue, we propose an efficient DM for IR (DiffIR), which consists of a compact IR prior extraction network (CPEN), dynamic IR transformer (DIRformer), and denoising network. Specifically, DiffIR has two training stages: pretraining and training DM. In pretraining, we input ground-truth images into CPENS1to capture a compact IR prior representation (IPR) to guide DIRformer. In the second stage, we train the DM to directly estimate the same IRP as pretrained CPENS1only using LQ images. We observe that since the IPR is only a compact vector, DiffIR can use fewer iterations than traditional DM to obtain accurate estimations and generate more stable and realistic results. Since the iterations are few, our DiffIR can adopt a joint optimization of CPENS2, DIRformer, and denoising network, which can further reduce the estimation error influence. We conduct extensive experiments on several IR tasks and achieve SOTA performance while consuming less computational costs. Code is available at https://github.com/Zj-BinXia/DiffIR.

Guided depth map super-resolution (GDSR), as a hot topic in multi-modal image processing, aims to upsample low-resolution (LR) depth maps with additional information involved in high-resolution (HR) RGB images from the same scene. The critical step of this task is to effectively extract domain-shared and domain-private RGB/depth features. In addition, three detailed issues, namely blurry edges, noisy surfaces, and over-transferred RGB texture, need to be addressed. In this paper, we propose the Spherical Space feature Decomposition Network (SSDNet) to solve the above issues. To better model cross-modality features, Restormer block-based RGB/depth encoders are employed for extracting local-global features. Then, the extracted features are mapped to the spherical space to complete the separation of private features and the alignment of shared features. Shared features of RGB are fused with the depth features to complete the GDSR task. Subsequently, a spherical contrast refinement (SCR) module is proposed to further address the detail issues. Patches that are classified according to imperfect categories are input into the SCR module, where the patch features are pulled closer to the ground truth and pushed away from the corresponding imperfect samples in the spherical feature space via contrastive learning. Extensive experiments demonstrate that our method can achieve state-of-the-art results on four test datasets, as well as successfully generalize to real-world scenes. The code is available at https://github.com/Zhaozixiang1228/GDSR-SSDNet.

Multi-modality image fusion aims to combine different modalities to produce fused images that retain the complementary features of each modality, such as functional highlights and texture details. To leverage strong generative priors and address challenges such as unstable training and lack of interpretability for GAN-based generative methods, we propose a novel fusion algorithm based on the denoising diffusion probabilistic model (DDPM). The fusion task is formulated as a conditional generation problem under the DDPM sampling framework, which is further divided into an unconditional generation subproblem and a maximum likelihood subproblem. The latter is modeled in a hierarchical Bayesian manner with latent variables and inferred by the expectation-maximization (EM) algorithm. By integrating the inference solution into the diffusion sampling iteration, our method can generate high-quality fused images with natural image generative priors and cross-modality information from source images. Note that all we required is an unconditional pre-trained generative model, and no fine-tuning is needed. Our extensive experiments indicate that our approach yields promising fusion results in infrared-visible image fusion and medical image fusion. The code is available at https://github.com/Zhaozixiang1228/MMIF-DDFM.

We introduce token-consistent dropout in vision transformers, which improves network calibration without causing any severe drop in performance. We use linear layers with token-consistent stochastic parameters inside the multilayer perceptron blocks, without altering the architecture of the transformer. The stochastic parameters are sampled from the uniform distribution, both during training and inference. The applied linear operations preserve the topological structure, formed by the set of tokens passing through the shared multilayer perceptron. This operation encourages the learning of the recognition task to rely on the topological structures of the tokens, instead of their values, which in turn offers the desired behavior. We compare our method to established baselines and applicable state-of-the-art alternatives, to demonstrate its impact on the calibration of vision transformers.

Activation functions provide the non-linearity to deep neural networks, which are crucial for the optimization and performance improvement. In this paper, we propose a learnable continuous piece-wise nonlinear activation function (or CPN in short), which improves the widely used ReLU from three directions, i.e., finer pieces, non-linear terms and learnable parameterization. CPN is a continuous activation function with multiple pieces and incorporates non-linear terms in every interval. We give a general formulation of CPN and provide different implementations according to three key factors: whether the activation space is divided uniformly or not, whether the non-linear terms exist or not, and whether the activation function is continuous or not. We demonstrate the effectiveness of our method on image classification and single image super-resolution tasks by simply changing the activation function. For example, CPN improves 4.78% / 4.52% top-1 accuracy over ReLU on MobileNetV2_0.25 / MobileNetV2_0.35 for ImageNet classification and achieves better PSNR on several benchmarks for super-resolution. Our implementation is available at https://github.com/xc-G/CPN.

As neuromorphic technology is maturing, its application to robotics and autonomous vehicle systems has become an area of active research. In particular, event cameras have emerged as a compelling alternative to frame-based cameras in low-power and latency-demanding applications. To enable event cameras to operate alongside staple sensors like lidar in perception tasks, we propose a direct, temporally-decoupled extrinsic calibration method between event cameras and lidars. The high dynamic range, high temporal resolution, and low-latency operation of event cameras are exploited to directly register lidar laser returns, allowing information-based correlation methods to optimize for the 6- DoF extrinsic calibration between the two sensors. This paper presents the first direct calibration method between event cameras and lidars, removing dependencies on frame-based camera intermediaries and/or highly-accurate hand measurements. Code: https://github.com/kev-in-ta/12e

Data-driven simulation has become a favorable way to train and test autonomous driving algorithms. The idea of replacing the actual environment with a learned simulator has also been explored in model-based reinforcement learning in the context of world models. In this work, we show data-driven traffic simulation can be formulated as a world model. We present TrafficBots, a multi-agent policy built upon motion prediction and end-to-end driving, and based on TrafficBots we obtain a world model tailored for the planning module of autonomous vehicles. Existing data-driven traffic simulators are lacking configurability and scalability. To generate configurable behaviors, for each agent we introduce a destination as nav-igational information, and a time-invariant latent personality that specifies the behavioral style. To improve the scalability, we present a new scheme of positional encoding for angles, allowing all agents to share the same vectorized context and the use of an architecture based on dot-product attention. As a result, we can simulate all traffic participants seen in dense urban scenarios. Experiments on the Waymo open motion dataset show TrafficBots can simulate realistic multi-agent behaviors and achieve good performance on the motion prediction task.

An emerging field of sequential decision problems is safe Reinforcement Learning (RL), where the objective is to maximize the reward while obeying safety constraints. Being able to handle constraints is essential for deploying RL agents in real-world environments, where constraint violations can harm the agent and the environment. To this end, we propose a safe model-free RL algorithm with a novel multiplicative value function consisting of a safety critic and a reward critic. The safety critic predicts the probability of constraint violation and discounts the reward critic that only estimates constraint-free returns. By splitting responsibilities, we facilitate the learning task leading to increased sample efficiency. We integrate our approach into two popular RL algorithms, Proximal Policy Optimization and Soft Actor-Critic, and evaluate our method in four safety-focused environments, including classical RL benchmarks augmented with safety constraints and robot navigation tasks with images and raw Lidar scans as observations. Finally, we make the zero-shot sim-to-real transfer where a differential drive robot has to navigate through a cluttered room. Our code can be found at https://github.com/nikeke19/Safe-Mult-RL.

The aim of this paper is to study the influence of locality mechanisms in vision transformers. Transformers originated from machine translation and are particularly good at modelling long-range dependencies within a long sequence. Although the global interaction between the token embeddings could be well modelled by the self-attention mechanism of transformers, what is lacking is a locality mechanism for infor-mation exchange within a local region. In this paper, locality mechanism is systematically investigated by carefully designed controlled experiments. We add locality to vision transformers into the feed-forward network. This seemingly simple solution is inspired by the comparison between feed-forward networks and inverted residual blocks. The importance of locality mechanisms is validated in two ways: 1) A wide range of design choices (activation function, layer placement, expansion ratio) are available for incorporating locality mechanisms and proper choices can lead to a performance gain over the baseline, and 2) The same locality mechanism is successfully applied to vision transformers with different architecture designs, which shows the generalization of the locality concept. For ImageNet2012 classification, the locality-enhanced transformers outperform the baselines Swin-T [1], DeiT-T [2] and PVT-T [3] by 1.0%, 2.6 % and 3.1 % with a negligible increase in the number of parameters and computational effort. Code is available at https://github.com/ofsoundof/LocalViT.

3D motion transfer aims at transferring the motion from a dynamic input sequence to a static 3D object and outputs an identical motion of the target with high-fidelity and realistic visual effects. In this work, we propose a novel 3D Transformer framework called LART for 3D motion transfer. With carefully-designed architectures, LART is able to implicitly learn the correspondence via a flexible geometry perception. Thus, unlike other existing methods, LART does not require any key point annotations or pre-defined correspondence between the motion source and target meshes and can also handle large-size full-detailed unseen 3D targets. Besides, we introduce a novel latent metric regularization on the Transformer for better motion generation. Our rationale lies in the observation that the decoded motions can be approximately expressed as linearly geometric distortion at the frame level. The metric preservation of motions could be translated to the formation of linear paths in the underlying latent space as a rigorous constraint to control the synthetic motions occurring in the construction of the latent space. The proposed LART shows a high learning efficiency with the need for a few samples from the AMASS dataset to generate motions with plausible visual effects. The experimental results verify the potential of our generative model in applications of motion transfer, content generation, temporal interpolation, and motion denoising. The code is made available: https://github.com/mikecheninoulu/LART.

Diffusion-based methods have shown impressive visual results in the text-to-image domain. They first learn a latent space using an autoencoder, then run a denoising process on the bottleneck to generate new samples. However, learning an autoencoder requires substantial data in the target domain. Such data is scarce for 3D generation, prohibiting the learning of large-scale diffusion models for 3D synthesis. We present a novel approach to the generation of static and articulated 3D assets that has a 3D autodecoder at its core. The 3D autodecoder framework embeds properties learned from the target dataset in the latent space, which can then be decoded into a volumetric representation for rendering view-consistent appearance and geometry. We then identify the appropriate intermediate volumetric latent space, and introduce robust normalization and de-normalization operations to learn a 3D diffusion from 2D images or monocular videos of rigid or articulated objects. Our approach is flexible enough to use either existing camera supervision or no camera information at all -- instead efficiently learning it during training. Our evaluations demonstrate that our generation results outperform state-of-the-art alternatives on various benchmark datasets and metrics, including multi-view image datasets of synthetic objects, real in-the-wild videos of moving people, and a large-scale, real video dataset of static objects.

Semantic segmentation datasets often exhibit two types of imbalance: \textit{class imbalance}, where some classes appear more frequently than others and \textit{size imbalance}, where some objects occupy more pixels than others. This causes traditional evaluation metrics to be biased towards \textit{majority classes} (e.g. overall pixel-wise accuracy) and \textit{large objects} (e.g. mean pixel-wise accuracy and per-dataset mean intersection over union). To address these shortcomings, we propose the use of fine-grained mIoUs along with corresponding worst-case metrics, thereby offering a more holistic evaluation of segmentation techniques. These fine-grained metrics offer less bias towards large objects, richer statistical information, and valuable insights into model and dataset auditing. Furthermore, we undertake an extensive benchmark study, where we train and evaluate 15 modern neural networks with the proposed metrics on 12 diverse natural and aerial segmentation datasets. Our benchmark study highlights the necessity of not basing evaluations on a single metric and confirms that fine-grained mIoUs reduce the bias towards large objects. Moreover, we identify the crucial role played by architecture designs and loss functions, which lead to best practices in optimizing fine-grained metrics. The code is available at \href{https://github.com/zifuwanggg/JDTLosses}{https://github.com/zifuwanggg/JDTLosses}.

The real-world deployment of an autonomous driving system requires its components to run on-board and in real-time, including the motion prediction module that predicts the future trajectories of surrounding traffic participants. Existing agent-centric methods have demonstrated outstanding performance on public benchmarks. However, they suffer from high computational overhead and poor scalability as the number of agents to be predicted increases. To address this problem, we introduce the K-nearest neighbor attention with relative pose encoding (KNARPE), a novel attention mechanism allowing the pairwise-relative representation to be used by Transformers. Then, based on KNARPE we present the Heterogeneous Polyline Transformer with Relative pose encoding (HPTR), a hierarchical framework enabling asynchronous token update during the online inference. By sharing contexts among agents and reusing the unchanged contexts, our approach is as efficient as scene-centric methods, while performing on par with state-of-the-art agent-centric methods. Experiments on Waymo and Argoverse-2 datasets show that HPTR achieves superior performance among end-to-end methods that do not apply expensive post-processing or model ensembling. The code is available at https://github.com/zhejz/HPTR.

The design of more complex and powerful neural network models has significantly advanced the state-of-the-art in visual object tracking. These advances can be attributed to deeper networks, or the introduction of new building blocks, such as transformers. However, in the pursuit of increased tracking performance, runtime is often hindered. Furthermore, efficient tracking architectures have received surprisingly little attention. In this paper, we introduce the Exemplar Transformer, a transformer module utilizing a single instance level attention layer for realtime visual object tracking. E.T.Track, our visual tracker that incorporates Exemplar Transformer modules, runs at 47 FPS on a CPU. This is up to 8× faster than other transformer-based models. When compared to lightweight trackers that can operate in realtime on standard CPUs, E.T.Track consistently outperforms all other methods on the LaSOT [16], OTB-100 [52], NFS [27], TrackingNet [36], and VOTST2020 [29] datasets. Code and models are available at https://github.com/pblatter/ettrack.

Due to the scarcity of dense pixel-level semantic annotations for images recorded in adverse visual conditions, there has been a keen interest in unsupervised domain adaptation (UDA) for the semantic segmentation of such images. UDA adapts models trained on normal conditions to the target adverse-condition domains. Meanwhile, multiple datasets with driving scenes provide corresponding images of the same scenes across multiple conditions, which can serve as a form of weak supervision for domain adaptation. We propose Refign, a generic extension to self-training-based UDA methods which leverages these cross-domain correspondences. Refign consists of two steps: (1) aligning the normal-condition image to the corresponding adverse-condition image using an uncertainty-aware dense matching network, and (2) refining the adverse prediction with the normal prediction using an adaptive label correction mechanism. We design custom modules to streamline both steps and set the new state of the art for domain-adaptive semantic segmentation on several adverse-condition benchmarks, including ACDC and Dark Zurich. The approach introduces no extra training parameters, minimal computational overhead—during training only—and can be used as a drop-in extension to improve any given self-training-based UDA method. Code is available at https://github.com/brdav/refign.

Video super-resolution (VSR) has many applications that pose strict causal, real-time, and latency constraints, including video streaming and TV. We address the VSR problem under these settings, which poses additional important challenges since information from future frames is unavailable. Importantly, designing efficient, yet effective frame alignment and fusion modules remain central problems. In this work, we propose a recurrent VSR architecture based on a deformable attention pyramid (DAP). Our DAP aligns and integrates information from the recurrent state into the current frame prediction. To circumvent the computational cost of traditional attention-based methods, we only attend to a limited number of spatial locations, which are dynamically predicted by the DAP. Comprehensive experiments and analysis of the proposed key innovations show the effectiveness of our approach. We significantly reduce processing time and computational complexity in comparison to state-of-the-art methods, while maintaining a high performance. We surpass state-of-the-art method EDVR-M on two standard benchmarks with a speed-up of over 3×.

Visual question answering is a vision-and-language multimodal task, that aims at predicting answers given samples from the question and image modalities. Most recent methods focus on learning a good joint embedding space of images and questions, either by improving the interaction between these two modalities, or by making it a more discriminant space. However, how informative this joint space is, has not been well explored. In this paper, we propose a novel regularization for VQA models, Constrained Optimization using Barlow’s theory (COB), that improves the information content of the joint space by minimizing the redundancy. It reduces the correlation between the learned feature components and thereby disentangles semantic concepts. Our model also aligns the joint space with the answer embedding space, where we consider the answer and image+question as two different ‘views’ of what in essence is the same semantic information. We propose a constrained optimization policy to balance the categorical and redundancy minimization forces. When built on the state-of-the-art GGE model, the resulting model improves VQA accuracy by 1.4% and 4% on the VQA-CP v2 and VQA v2 datasets respectively. The model also exhibits better interpretability. Code is made available: https://github.com/abskjha/Barlow-constrained-VQA

Multi-task learning promises better model generalization on a target task by jointly optimizing it with an auxiliary task. However, the current practice requires additional labeling efforts for the auxiliary task, while not guaranteeing better model performance. In this paper, we find that jointly training a dense prediction (target) task with a self-supervised (auxiliary) task can consistently improve the performance of the target task, while eliminating the need for labeling auxiliary tasks. We refer to this joint training as Composite Learning (CompL). Experiments of CompL on monocular depth estimation, semantic segmentation, and boundary detection show consistent performance improvements in fully and partially labeled datasets. Further analysis on depth estimation reveals that joint training with self-supervision outperforms most labeled auxiliary tasks. We also find that CompL can improve model robustness when the models are evaluated in new domains. These results demonstrate the benefits of self-supervision as an auxiliary task, and establish the design of novel task-specific self-supervised methods as a new axis of investigation for future multi-task learning research.

Multi-view photometric stereo (MVPS) is a preferred method for detailed and precise 3D acquisition of an object from images. Although popular methods for MVPS can provide outstanding results, they are often complex to execute and limited to isotropic material objects. To address such limitations, we present a simple, practical approach to MVPS, which works well for isotropic as well as other object material types such as anisotropic and glossy. The proposed approach in this paper exploits the benefit of uncertainty modeling in a deep neural network for a reliable fusion of photometric stereo (PS) and multi-view stereo (MVS) network predictions. Yet, contrary to the recently proposed state-of-the-art, we introduce neural volume rendering methodology for a trustworthy fusion of MVS and PS measurements. The advantage of introducing neural volume rendering is that it helps in the reliable modeling of objects with diverse material types, where existing MVS methods, PS methods, or both may fail. Furthermore, it allows us to work on neural 3D shape representation, which has recently shown outstanding results for many geometric processing tasks. Our suggested new loss function aims to fit the zero level set of the implicit neural function using the most certain MVS and PS network predictions coupled with weighted neural volume rendering cost. The proposed approach shows state-of-the-art results when tested extensively on several benchmark datasets.

Compositional zero-shot learning aims to recognize unseen compositions of seen visual primitives of object classes and their states. While all primitives (states and objects) are observable during training in some combination, their complex interaction makes this task especially hard. For example, wet changes the visual appearance of a dog very differently from a bicycle. Furthermore, we argue that relationships between compositions go beyond shared states or objects. A cluttered office can contain a busy table; even though these compositions don’t share a state or object, the presence of a busy table can guide the presence of a cluttered office. We propose a novel method called Compositional Attention Propagated Embedding (CAPE) as a solution. The key intuition to our method is that a rich dependency structure exists between compositions arising from complex interactions of primitives in addition to other dependencies between compositions. CAPE learns to identify this structure and propagates knowledge between them to learn class embedding for all seen and unseen compositions. In the challenging generalized compositional zero-shot setting, we show that our method outperforms previous baselines to set a new state-of-the-art on three publicly available benchmarks.

A single-shot multispectral camera equipped with an optimized color filter array (CFA) has the potential to deliver a fast and low-cost hyperspectral (HS) imaging system. Previous solutions are largely restricted to designing demosaicing algorithms for fixed CFAs – be it the Bayer color pattern or evenly-spaced spectral multiplexing patterns. Since sampling and reconstruction are tightly-coupled, the ability to search for an optimal solution is severely constrained by using predefined CFAs. In this work, we simultaneously address the problem of spectral band selection, CFA design, image demosaicing, and spectral image recovery in a joint learning framework for single-shot HS imaging. We propose a reinforcement learning (RL) based method for spectral band selection and a novel neural network for CFA generation, image demosaicing, and HS image recovery. The final spectral reconstruction accuracy is used to supervise the training of the main network to maximize the synergies between those tightly-related tasks. The RL method regards the main network as an agent to collect reward. Our final method delivers a simple setup – as simple as an RGB camera – for HS imaging. Experimental results show that our method outperforms competing methods by a large margin.

There have been emerging a number of benchmarks and techniques for the detection of deepfakes. However, very few works study the detection of incrementally appearing deepfakes in the real-world scenarios. To simulate the wild scenes, this paper suggests a continual deepfake detection benchmark (CDDB) over a new collection of deepfakes from both known and unknown generative models. The suggested CDDB designs multiple evaluations on the detection over easy, hard, and long sequence of deepfake tasks, with a set of appropriate measures. In addition, we exploit multiple approaches to adapt multiclass incremental learning methods, commonly used in the continual visual recognition, to the continual deepfake detection problem. We evaluate existing methods, including their adapted ones, on the proposed CDDB. Within the proposed benchmark, we explore some commonly known essentials of standard continual learning. Our study provides new insights on these essentials in the context of continual deepfake detection. The suggested CDDB is clearly more challenging than the existing benchmarks, which thus offers a suitable evaluation avenue to the future research. Both data and code are available at https://github.com/Coral79/CDDB.

We propose a novel domain adaptive action detection approach and a new adaptation protocol that leverages the recent advancements in image-level unsupervised domain adaptation (UDA) techniques and handle vagaries of instance-level video data. Self-training combined with cross-domain mixed sampling has shown remarkable performance gain in semantic segmentation in UDA (unsupervised domain adaptation) context. Motivated by this fact, we propose an approach for human action detection in videos that transfers knowledge from the source domain (annotated dataset) to the target domain (unannotated dataset) using mixed sampling and pseudo-label-based self-training. The existing UDA techniques follow a Class-Mix algorithm for semantic segmentation. However, simply adopting ClassMix for action detection does not work, mainly because these are two entirely different problems, i.e., pixel-label classification vs. instance-label detection. To tackle this, we propose a novel action instance mixed sampling technique that combines information across domains based on action instances instead of action classes. Moreover, we propose a new UDA training protocol that addresses the long-tail sample distribution and domain shift problem by using supervision from an auxiliary source domain (ASD). For the ASD, we propose a new action detection dataset with dense frame-level annotations. We name our proposed framework as domain-adaptive action instance mixing (DA-AIM). We demonstrate that DA-AIM consistently outperforms prior works on challenging domain adaptation benchmarks. The source code is available at https://github.com/wwwfan628/DA-AIM.

In most scenarios, conditional image generation can be thought of as an inversion of the image understanding process. Since generic image understanding involves solving multiple tasks, it is natural to aim at generating images via multi-conditioning. However, multi-conditional image generation is a very challenging problem due to the heterogeneity and the sparsity of the (in practice) available conditioning labels. In this work, we propose a novel neural architecture to address the problem of heterogeneity and sparsity of the spatially multi-conditional labels. Our choice of spatial conditioning, such as by semantics and depth, is driven by the promise it holds for better control of the image generation process. The proposed method uses a transformer-like architecture operating pixel-wise, which receives the available labels as input tokens to merge them in a learned homogeneous space of labels. The merged labels are then used for image generation via conditional generative adversarial training. In this process, the sparsity of the labels is handled by simply dropping the input tokens corresponding to the missing labels at the desired locations, thanks to the proposed pixel-wise operating architecture. Our experiments on three benchmark datasets demonstrate the clear superiority of our method over the state-of-the-art and compared baselines. The source code can be found at https://github.com/96ritika/TLAM.

Current methods for spatio-temporal action tube detection often extend a bounding box proposal at a given key-frame into a 3D temporal cuboid and pool features from nearby frames. However, such pooling fails to accumulate meaningful spatio-temporal features if the position or shape of the actor shows large 2D motion and variability through the frames, due to large camera motion, large actor shape deformation, fast actor action and so on. In this work, we aim to study the performance of cuboid-aware feature aggregation in action detection under large action. Further, we propose to enhance actor feature representation under large motion by tracking actors and performing temporal feature aggregation along the respective tracks. We define the actor motion with intersection-over-union (IoU) between the boxes of action tubes/tracks at various fixed time scales. The action having a large motion would result in lower IoU over time, and slower actions would maintain higher IoU. We find that track-aware feature aggregation consistently achieves a large improvement in action detection performance, especially for actions under large motion compared to cuboid-aware baseline. As a result, we also report state-of-the-art on the large-scale MultiSports dataset.

Abstract Unsupervised template discovery via implicit representation in a category of shapes has recently shown strong performance. At the core, such methods deform input shapes to a common template space which allows establishing correspondences as well as implicit representation of the shapes. In this work we investigate the inherent assumption that the implicit neural field optimization naturally leads to consistently warped shapes, thus providing both good shape reconstruction and correspondences. Contrary to this convenient assumption, in practice we observe that such is not the case, consequently resulting in sub‐optimal point correspondences. In order to solve the problem, we re‐visit the warp design and more importantly introduce explicit constraints using unsupervised sparse point predictions, directly encouraging consistency of the warped shapes. We use the unsupervised sparse keypoints in order to further condition the deformation warp and enforce the consistency of the deformation warp. Experiments in dynamic non‐rigid DFaust and ShapeNet categories show that our problem identification and solution provide the new state‐of‐the‐art in unsupervised dense correspondences.

Abstract Video enhancement is a challenging problem, more than that of stills, mainly due to high computational cost, larger data volumes and the difficulty of achieving consistency in the spatio-temporal domain. In practice, these challenges are often coupled with the lack of example pairs, which inhibits the application of supervised learning strategies. To address these challenges, we propose an efficient adversarial video enhancement framework that learns directly from unpaired video examples. In particular, our framework introduces new recurrent cells that consist of interleaved local and global modules for implicit integration of spatial and temporal information. The proposed design allows our recurrent cells to efficiently propagate spatio-temporal information across frames and reduces the need for high complexity networks. Our setting enables learning from unpaired videos in a cyclic adversarial manner, where the proposed recurrent units are employed in all architectures. Efficient training is accomplished by introducing one single discriminator that learns the joint distribution of source and target domain simultaneously. The enhancement results demonstrate clear superiority of the proposed video enhancer over the state-of-the-art methods, in all terms of visual quality, quantitative metrics, and inference speed. Notably, our video enhancer is capable of enhancing over 35 frames per second of FullHD video (1080x1920).

Abstract Training deep networks for semantic segmentation requires large amounts of labeled training data, which presents a major challenge in practice, as labeling segmentation masks is a highly labor-intensive process. To address this issue, we present a framework for semi-supervised and domain-adaptive semantic segmentation, which is enhanced by self-supervised monocular depth estimation (SDE) trained only on unlabeled image sequences. In particular, we utilize SDE as an auxiliary task comprehensively across the entire learning framework: First, we automatically select the most useful samples to be annotated for semantic segmentation based on the correlation of sample diversity and difficulty between SDE and semantic segmentation. Second, we implement a strong data augmentation by mixing images and labels using the geometry of the scene. Third, we transfer knowledge from features learned during SDE to semantic segmentation by means of transfer and multi-task learning. And fourth, we exploit additional labeled synthetic data with Cross-Domain DepthMix and Matching Geometry Sampling to align synthetic and real data. We validate the proposed model on the Cityscapes dataset, where all four contributions demonstrate significant performance gains, and achieve state-of-the-art results for semi-supervised semantic segmentation as well as for semi-supervised domain adaptation. In particular, with only 1/30 of the Cityscapes labels, our method achieves 92% of the fully-supervised baseline performance and even 97% when exploiting additional data from GTA. The source code is available at https://github.com/lhoyer/improving_segmentation_with_selfsupervised_depth .

Humans can robustly recognize and localize objects by using visual and/or auditory cues. While machines are able to do the same with visual data already, less work has been done with sounds. This work develops an approach for scene understanding purely based on binaural sounds. The considered tasks include predicting the semantic masks of sound-making objects, the motion of sound-making objects, and the depth map of the scene. To this aim, we propose a novel sensor setup and record a new audio-visual dataset of street scenes with eight professional binaural microphones and a 360$\mathrm{^{\circ }}$camera. The co-existence of visual and audio cues is leveraged for supervision transfer. In particular, we employ a cross-modal distillation framework that consists of multiple vision ‘teacher’ methods and a sound ‘student’ method – the student method is trained to generate the same results as the teacher methods do. This way, the auditory system can be trained without using human annotations. To further boost the performance, we propose another novel auxiliary task, coined Spatial Sound Super-Resolution, to increase the directional resolution of sounds. We then formulate the four tasks into one end-to-end trainable multi-tasking network aiming to boost the overall performance. Experimental results show that 1) our method achieves good results for all four tasks, 2) the four tasks are mutually beneficial – training them together achieves the best performance, 3) the number and orientation of microphones are both important, and 4) features learned from the standard spectrogram and features obtained by the classic signal processing pipeline are complementary for auditory perception tasks. The data and code are released on the project page:https://www.trace.ethz.ch/publications/2020/sound_perception/index.html.

We propose a noveledge guidedgenerativeadversarialnetwork withcontrastive learning (ECGAN) for the challenging semantic image synthesis task. Although considerable improvements have been achieved by the community in the recent period, the quality of synthesized images is far from satisfactory due to three largely unresolved challenges. 1) The semantic labels do not provide detailed structural information, making it challenging to synthesize local details and structures; 2) The widely adopted CNN operations such as convolution, down-sampling, and normalization usually cause spatial resolution loss and thus cannot fully preserve the original semantic information, leading to semantically inconsistent results (e.g., missing small objects); 3) Existing semantic image synthesis methods focus on modeling “local” semantic information from a single input semantic layout. However, they ignore “global” semantic information of multiple input semantic layouts, i.e., semantic cross-relations between pixels across different input layouts. To tackle 1), we propose to use the edge as an intermediate representation which is further adopted to guide image generation via a proposed attention guided edge transfer module. Edge information is produced by a convolutional generator and introduces detailed structure information. To tackle 2), we design an effective module to selectively highlight class-dependent feature maps according to the original semantic layout to preserve the semantic information. To tackle 3), inspired by current methods in contrastive learning, we propose a novel contrastive learning method, which aims to enforce pixel embeddings belonging to the same semantic class to generate more similar image content than those from different classes. We further propose a novel multi-scale contrastive learning method that aims to push same-class features from different scales closer together being able to capture more semantic relations by explicitly exploring the structures of labeled pixels from multiple input semantic layouts from different scales. Experiments on three challenging datasets show that our methods achieve significantly better results than state-of-the-art approaches. The source code is available athttps://github.com/Ha0Tang/ECGAN.

Establishing robust and accurate correspondences between a pair of images is a long-standing computer vision problem with numerous applications. While classically dominated by sparse methods, emerging dense approaches offer a compelling alternative paradigm that avoids the keypoint detection step. However, dense flow estimation is often inaccurate in the case of large displacements, occlusions, or homogeneous regions. In order to apply dense methods to real-world applications, such as pose estimation, image manipulation, or 3D reconstruction, it is therefore crucial to estimate the confidence of the predicted matches. We propose the Enhanced Probabilistic Dense Correspondence Network, PDC-Net+, capable of estimating accurate dense correspondences along with a reliable confidence map. We develop a flexible probabilistic approach that jointly learns the flow prediction and its uncertainty. In particular, we parametrize the predictive distribution as a constrained mixture model, ensuring better modelling of both accurate flow predictions and outliers. Moreover, we develop an architecture and an enhanced training strategy tailored for robust and generalizable uncertainty prediction in the context of self-supervised training. Our approach obtains state-of-the-art results on multiple challenging geometric matching and optical flow datasets. We further validate the usefulness of our probabilistic confidence estimation for the tasks of pose estimation, 3D reconstruction, image-based localization, and image retrieval.

Efficient image super-resolution (SR) has witnessed rapid progress thanks to novel lightweight architectures or model compression techniques (e.g., neural architecture search and knowledge distillation). Nevertheless, these methods consume considerable resources or/and neglect to squeeze out the network redundancy at a more fine-grained convolution filter level. Network pruning is a promising alternative to overcome these shortcomings. However, structured pruning is known to be tricky when applied to SR networks because the extensive residual blocks demand the pruned indices of different layers to be the same. Besides, the principled determination of proper layerwise sparsities remains challenging too. In this article, we present Global Aligned Structured Sparsity Learning (GASSL) to resolve these problems. GASSL has two major components: Hessian-Aided Regularization (HAIR) and Aligned Structured Sparsity Learning (ASSL). HAIR is a regularization-based sparsity auto-selection algorithm with Hessian considered implicitly. A proven proposition is introduced to justify its design. ASSL is for physically pruning SR networks. Particularly, a new penalty term Sparsity Structure Alignment (SSA) is proposed to align the pruned indices of different layers. With GASSL, we design two new efficient single image SR networks of different architecture genres, pushing the efficiency envelope of SR models one step forward. Extensive results demonstrate the merits of GASSL over other recent counterparts.

In this paper, we present a novel end-to-end group collaborative learning network, termed GCoNet+, which can effectively and efficiently (250 fps) identify co-salient objects in natural scenes. The proposed GCoNet+ achieves the new state-of-the-art performance for co-salient object detection (CoSOD) through mining consensus representations based on the following two essential criteria: 1) intra-group compactness to better formulate the consistency among co-salient objects by capturing their inherent shared attributes using our novel group affinity module (GAM); 2) inter-group separability to effectively suppress the influence of noisy objects on the output by introducing our new group collaborating module (GCM) conditioning on the inconsistent consensus. To further improve the accuracy, we design a series of simple yet effective components as follows: i) a recurrent auxiliary classification module (RACM) promoting model learning at the semantic level; ii) a confidence enhancement module (CEM) assisting the model in improving the quality of the final predictions; and iii) a group-based symmetric triplet (GST) loss guiding the model to learn more discriminative features. Extensive experiments on three challenging benchmarks, i.e., CoCA, CoSOD3k, and CoSal2015, demonstrate that our GCoNet+ outperforms the existing 12 cutting-edge models. Code has been released at https://github.com/ZhengPeng7/GCoNet_plus.

Video segmentation-partitioning video frames into multiple segments or objects-plays a critical role in a broad range of practical applications, from enhancing visual effects in movie, to understanding scenes in autonomous driving, to creating virtual background in video conferencing. Recently, with the renaissance of connectionism in computer vision, there has been an influx of deep learning based approaches for video segmentation that have delivered compelling performance. In this survey, we comprehensively review two basic lines of research - generic object segmentation (of unknown categories) in videos, and video semantic segmentation - by introducing their respective task settings, background concepts, perceived need, development history, and main challenges. We also offer a detailed overview of representative literature on both methods and datasets. We further benchmark the reviewed methods on several well-known datasets. Finally, we point out open issues in this field, and suggest opportunities for further research. We also provide a public website to continuously track developments in this fast advancing field: https://github.com/tfzhou/VS-Survey.

Recent years have witnessed an increasing interest in end-to-end learned video compression. Most previous works explore temporal redundancy by detecting and compressing a motion map to warp the reference frame towards the target frame. Yet, it failed to adequately take advantage of the historical priors in the sequential reference frames. In this paper, we propose an Advanced Learned Video Compression (ALVC) approach with the in-loop frame prediction module, which is able to effectively predict the target frame from the previously compressed frames, without consuming any bit-rate. The predicted frame can serve as a better reference than the previously compressed frame, and therefore it benefits the compression performance. The proposed in-loop prediction module is a part of the end-to-end video compression and is jointly optimized in the whole framework. We propose the recurrent and the bi-directional in-loop prediction modules for compressing P-frames and B-frames, respectively. The experiments show the state-of-the-art performance of our ALVC approach in learned video compression. We also outperform the default hierarchical B mode of x265 in terms of PSNR and beat the slowest mode of the SSIM-tuned x265 on MS-SSIM. The project page:https://github.com/RenYang-home/ALVC.

Normalizing Flows (NFs) are flexible explicit generative models that have been shown to accurately model complex real-world data distributions. However, their invertibility constraint imposes limitations on data distributions that reside on lower dimensional manifolds embedded in higher dimensional space. Practically, this shortcoming is often bypassed by adding noise to the data which impacts the quality of the generated samples. In contrast to prior work, we approach this problem by generating samples from the original data distribution given full knowledge about the perturbed distribution and the noise model. To this end, we establish that NFs trained on perturbed data implicitly represent the manifold in regions of maximum likelihood. Then, we propose an optimization objective that recovers the most likely point on the manifold given a sample from the perturbed distribution. Finally, we focus on 3D point clouds for which we utilize the explicit nature of NFs, i.e. surface normals extracted from the gradient of the log-likelihood and the log-likelihood itself, to apply Poisson surface re-construction to refine generated point sets.

Optimization based tracking methods have been widely successful by integrating a target model prediction module, providing effective global reasoning by minimizing an objective function. While this inductive bias integrates valuable domain knowledge, it limits the expressivity of the tracking network. In this work, we therefore propose a tracker architecture employing a Transformer-based model prediction module. Transformers capture global relations with little inductive bias, allowing it to learn the prediction of more powerful target models. We further extend the model predictor to estimate a second set of weights that are applied for accurate bounding box regression. The resulting tracker ToMP relies on training and on test frame information in order to predict all weights transductively. We train the proposed tracker end-to-end and validate its performance by conducting comprehensive experiments on multiple tracking datasets. ToMP sets a new state of the art on three benchmarks, achieving an AUC of 68.5% on the challenging LaSOT [14] dataset. The code and trained models are available at https://github.com/visionml/pytracking

Test-time domain adaptation aims to adapt a source pre-trained model to a target domain without using any source data. Existing works mainly consider the case where the target domain is static. However, real-world machine perception systems are running in non-stationary and continually changing environments where the target domain distribution can change over time. Existing methods, which are mostly based on self-training and entropy regularization, can suffer from these non-stationary environments. Due to the distribution shift over time in the target domain, pseudo-labels become unreliable. The noisy pseudo-labels can further lead to error accumulation and catastrophic forgetting. To tackle these issues, we propose a continual test-time adaptation approach (CoTTA) which comprises two parts. Firstly, we propose to reduce the error accumulation by using weight-averaged and augmentation-averaged predictions which are often more accurate. On the other hand, to avoid catastrophic forgetting, we propose to stochastically restore a small part of the neurons to the source pre-trained weights during each iteration to help preserve source knowledge in the longterm. The proposed method enables the longterm adaptation for all parameters in the network. CoTTA is easy to implement and can be readily incorporated in off-the-shelf pre-trained models. We demonstrate the effectiveness of our approach on four classification tasks and a segmentation task for continual test-time adaptation, on which we outperform existing methods. Our code is available at https://gin.ee/cotta.

Hyperspectral image (HSI) reconstruction aims to recover the 3D spatial-spectral signal from a 2D measurement in the coded aperture snapshot spectral imaging (CASSI) system. The HSI representations are highly similar and correlated across the spectral dimension. Modeling the inter-spectra interactions is beneficial for HSI reconstruction. However, existing CNN-based methods show limitations in capturing spectral-wise similarity and long-range dependencies. Besides, the HSI information is modulated by a coded aperture (physical mask) in CASSI. Nonetheless, current algorithms have not fully explored the guidance effect of the mask for HSI restoration. In this paper, we propose a novel framework, Mask-guided Spectral-wise Transformer (MST), for HSI reconstruction. Specifically, we present a Spectral-wise Multi-head Self-Attention (S-MSA) that treats each spectral feature as a token and calculates self-attention along the spectral dimension. In addition, we customize a Mask-guided Mechanism (MM) that directs S- MSA to pay attention to spatial regions with high-fidelity spectral representations. Extensive experiments show that our MST significantly outperforms state-of-the-art (SOTA) methods on simulation and real HSI datasets while requiring dramatically cheaper computational and memory costs. https://github.com/caiyuanhao1998/MST/

We propose a pipeline to generate Neural Radiance Fields (NeRF) of an object or a scene of a specific class, conditioned on a single input image. This is a challenging task, as training NeRF requires multiple views of the same scene, coupled with corresponding poses, which are hard to obtain. Our method is based on$\pi$-GAN, a generative model for unconditional 3D-aware image synthesis, which maps random latent codes to radiance fields of a class of objects. We jointly optimize (1) the$\pi$-GAN objective to utilize its high-fidelity 3D-aware generation and (2) a carefully designed reconstruction objective. The latter includes an encoder coupled with$\pi$-GAN generator to form an autoencoder. Unlike previous few-shot NeRF approaches, our pipeline is unsupervised, capable of being trained with independent images without 3D, multi-view, or pose supervision. Applications of our pipeline include 3d avatar generation, object-centric novel view synthesis with a single input image, and 3d-aware super-resolution, to name a few.

Knowledge of the road network topology is crucial for autonomous planning and navigation. Yet, recovering such topology from a single image has only been explored in part. Furthermore, it needs to refer to the ground plane, where also the driving actions are taken. This paper aims at extracting the local road network topology, directly in the bird’ s-eye- view (BEV), all in a complex urban set-ting. The only input consists of a single onboard, for-ward looking camera image. We represent the road topology using a set of directed lane curves and their interactions, which are captured using their intersection points. To better capture topology, we introduce the concept of minimal cycles and their covers. A minimal cycle is the smallest cycle formed by the directed curve segments (be-tween two intersections). The cover is a set of curves whose segments are involved in forming a minimal cycle. We first show that the covers suffice to uniquely represent the road topology. The covers are then used to supervise deep neural networks, along with the lane curve supervision. These learn to predict the road topology from a single input image. The results on the NuScenes and Argo-verse benchmarks are significantly better than those ob-tained with baselines. Code: https://github.com/ybarancan/TopologicalLaneGraph.

3D object detection is a central task for applications such as autonomous driving, in which the system needs to localize and classify surrounding traffic agents, even in the presence of adverse weather. In this paper, we address the problem of LiDAR-based 3D object detection under snow-fall. Due to the difficulty of collecting and annotating training data in this setting, we propose a physically based method to simulate the effect of snowfall on real clear-weather LiDAR point clouds. Our method samples snow particles in 2D space for each LiDAR line and uses the in-duced geometry to modify the measurement for each LiDAR beam accordingly. Moreover, as snowfall often causes wet-ness on the ground, we also simulate ground wetness on LiDAR point clouds. We use our simulation to generate par-tially synthetic snowy LiDAR data and leverage these data for training 3D object detection models that are robust to snowfall. We conduct an extensive evaluation using several state-of-the-art 3D object detection methods and show that our simulation consistently yields significant performance gains on the real snowy STF dataset compared to clear-weather baselines and competing simulation approaches, while not sacrificing performance in clear weather. Our code is available at github.com/SysCV/LiDAR_snow_sim.

As acquiring pixel-wise annotations of real-world images for semantic segmentation is a costly process, a model can instead be trained with more accessible synthetic data and adapted to real images without requiring their annotations. This process is studied in unsupervised domain adaptation (UDA). Even though a large number of methods propose new adaptation strategies, they are mostly based on outdated network architectures. As the influence of recent network architectures has not been systematically studied, we first benchmark different network architectures for UDA and newly reveal the potential of Transformers for UDA semantic segmentation. Based on the findings, we propose a novel UDA method, DAFormer. The network architecture of DAFormer consists of a Transformer encoder and a multi-level context-aware feature fusion decoder. It is enabled by three simple but crucial training strategies to stabilize the training and to avoid overfitting to the source domain: While (1) Rare Class Sampling on the source domain improves the quality of the pseudo-labels by mitigating the confirmation bias of self-training toward common classes, (2) a Thing-Class ImageNet Feature Distance and (3) a learning rate warmup promote feature transfer from ImageNet pretraining. DAFormer represents a major advance in UDA. It improves the state of the art by 10.8 mIoU for GTA→Cityscapes and 5.4 mIoU for Syruhia→Cityscapes and enables learning even difficult classes such as train, bus, and truck well. The implementation is available at https://github.com/lhoyer/DAFormer.

The rapid development of deep learning provides a better solution for the end-to-end reconstruction of hyperspectral image (HSI). However, existing learning-based methods have two major defects. Firstly, networks with self-attention usually sacrifice internal resolution to balance model performance against complexity, losing fine-grained high-resolution (HR) features. Secondly, even if the optimization focusing on spatial-spectral domain learning (SDL) converges to the ideal solution, there is still a significant visual difference between the reconstructed HSI and the truth. So we propose a high-resolution dual-domain learning network (HDNet) for HSI reconstruction. On the one hand, the proposed HR spatial-spectral attention module with its efficient feature fusion provides continuous and fine pixel-level features. On the other hand, frequency domain learning (FDL) is introduced for HSI reconstruction to narrow the frequency domain discrepancy. Dynamic FDL supervision forces the model to reconstruct fine-grained frequencies and compensate for excessive smoothing and distortion caused by pixel-level losses. The HR pixel-level attention and frequency-level refinement in our HDNet mutually promote HSI perceptual quality. Extensive quantitative and qualitative experiments show that our method achieves SOTA performance on simulated and real HSI datasets. https://github.com/Huxiaowan/HDNet

This paper presents a simple and effective solution to the longstanding classical multi-view photometric stereo (MVPS) problem. It is well-known that photometric stereo (PS) is excellent at recovering high-frequency surface details, whereas multi-view stereo (MVS) can help remove the low-frequency distortion due to PS and retain the global geometry of the shape. This paper proposes an approach that can effectively utilize such complementary strengths of PS and MVS. Our key idea is to combine them suitably while considering the per-pixel uncertainty of their estimates. To this end, we estimate per-pixel surface normals and depth using an uncertainty-aware deep-PS network and deep-MVS network, respectively. Uncertainty modeling helps select reliable surface normal and depth estimates at each pixel which then act as a true representative of the dense surface geometry. At each pixel, our approach either selects or discards deep-PS and deep-MVS network prediction depending on the prediction uncertainty measure. For dense, detailed, and precise inference of the object's surface profile, we propose to learn the implicit neural shape representation via a multilayer perceptron (MLP). Our approach encourages the MLP to converge to a natural zero-level set surface using the confident prediction from deep-PS and deep-MVS networks, providing superior dense surface reconstruction. Extensive experiments on the DiLiGenT-MV benchmark dataset show that our method provides high-quality shape recovery with a much lower memory footprint while outperforming almost all of the existing approaches.

Estimating 3D human poses from monocular videos is a challenging task due to depth ambiguity and self-occlusion. Most existing works attempt to solve both issues by exploiting spatial and temporal relationships. However, those works ignore the fact that it is an inverse problem where multiple feasible solutions (i.e., hypotheses) exist. To relieve this limitation, we propose a Multi-Hypothesis Transformer (MHFormer) that learns spatio-temporal representations of multiple plausible pose hypotheses. In order to effectively model multi-hypothesis dependencies and build strong relationships across hypothesis features, the task is decomposed into three stages: (i) Generate multiple initial hypothesis representations; (ii) Model self-hypothesis communication, merge multiple hypotheses into a single converged representation and then partition it into several diverged hypotheses; (iii) Learn cross-hypothesis communication and aggregate the multi-hypothesis features to synthesize the final 3D pose. Through the above processes, the final representation is enhanced and the synthesized pose is much more accurate. Extensive experiments show that MHFormer achieves state-of-the-art results on two challenging datasets: Human3.6M and MPI-INF-3DHP. Without bells and whistles, its performance surpasses the previous best result by a large margin of 3% on Human3.6M. Code and models are available at https://github.com/Vegetebird/MHFormer.

Channel (or 3D filter) pruning serves as an effective way to accelerate the inference of neural networks. There has been a flurry of algorithms that try to solve this practical problem, each being claimed effective in some ways. Yet, a benchmark to compare those algorithms directly is lacking, mainly due to the complexity of the algorithms and some custom settings such as the particular network configuration or training procedure. A fair benchmark is important for the further development of channel pruning. Meanwhile, recent investigations reveal that the channel configurations discovered by pruning algorithms are at least as important as the pre-trained weights. This gives channel pruning a new role, namely searching the optimal channel configuration. In this paper, we try to determine the channel configuration of the pruned models by random search. The proposed approach provides a new way to compare different methods, namely how well they behave compared with random pruning. We show that this simple strategy works quite well compared with other channel pruning methods. We also show that under this setting, there are surprisingly no clear winners among different channel importance evaluation methods, which then may tilt the research efforts into advanced channel configuration searching methods. Code will be released at https://github.com/ofsoundof/random_channel_pruning.

Free-form inpainting is the task of adding new content to an image in the regions specified by an arbitrary binary mask. Most existing approaches train for a certain distribution of masks, which limits their generalization capabilities to unseen mask types. Furthermore, training with pixel-wise and perceptual losses often leads to simple textural extensions towards the missing areas instead of semantically meaningful generation. In this work, we propose RePaint: A Denoising Diffusion Probabilistic Model (DDPM) based inpainting approach that is applicable to even extreme masks. We employ a pretrained unconditional DDPM as the generative prior. To condition the generation process, we only alter the reverse diffusion iterations by sampling the unmasked regions using the given image infor-mation. Since this technique does not modify or condition the original DDPM network itself, the model produces high-quality and diverse output images for any inpainting form. We validate our method for both faces and general-purpose image inpainting using standard and extreme masks. Re-Paint outperforms state-of-the-art Autoregressive, and GAN approaches for at least five out of six mask distributions. Github Repository: git.io/RePaint

Positional encodings have enabled recent works to train a single adversarial network that can generate images of different scales. However, these approaches are either limited to a set of discrete scales or struggle to maintain good perceptual quality at the scales for which the model is not trained explicitly. We propose the design of scale-consistent positional encodings invariant to our generator's layers transformations. This enables the generation of arbitrary-scale images even at scales unseen during training. Moreover, we incorporate novel inter-scale augmentations into our pipeline and partial generation training to facilitate the synthesis of consistent images at arbitrary scales. Lastly, we show competitive results for a continuum of scales on various commonly used datasets for image synthesis.

Monocular depth estimation is vital for scene understanding and downstream tasks. We focus on the supervised setup, in which ground-truth depth is available only at training time. Based on knowledge about the high regularity of real 3D scenes, we propose a method that learns to selectively leverage information from coplanar pixels to improve the predicted depth. In particular, we introduce a piecewise planarity prior which states that for each pixel, there is a seed pixel which shares the same planar 3D surface with the former. Motivated by this prior, we design a network with two heads. The first head outputs pixel-level plane coefficients, while the second one outputs a dense off-set vector field that identifies the positions of seed pixels. The plane coefficients of seed pixels are then used to predict depth at each position. The resulting prediction is adaptively fused with the initial prediction from the first head via a learned confidence to account for potential deviations from precise local planarity. The entire architecture is trained end-to-end thanks to the differentiability of the proposed modules and it learns to predict regular depth maps, with sharp edges at occlusion boundaries. An extensive evaluation of our method shows that we set the new state of the art in supervised monocular depth estimation, surpassing prior methods on NYU Depth-v2 and on the Garg split of KITTI. Our method delivers depth maps that yield plausible 3D reconstructions of the input scenes. Code is available at: https://github.com/SysCV/P3Depth

Flow-based generative models have shown an excellent ability to explicitly learn the probability density function of data via a sequence of invertible transformations. Yet, learning attentions in generative flows remains understudied, while it has made breakthroughs in other domains. To fill the gap, this paper introduces two types of invertible attention mechanisms, i.e., map-based and transformer-based attentions, for both unconditional and conditional generative flows. The key idea is to exploit a masked scheme of these two attentions to learn long-range data dependencies in the context of generative flows. The masked scheme allows for invertible attention modules with tractable Jacobian determinants, enabling its seamless integration at any positions of the flow-based models. The proposed attention mechanisms lead to more efficient generative flows, due to their capability of modeling the long-term data dependencies. Evaluation on multiple image synthesis tasks shows that the proposed attention flows result in efficient models and compare favorably against the state-of-the-art unconditional and conditional generative flows.

The contextual information plays a core role in semantic segmentation. As for video semantic segmentation, the contexts include static contexts and motional contexts, corresponding to static content and moving content in a video clip, respectively. The static contexts are well exploited in image semantic segmentation by learning multi-scale and global/long-range features. The motional contexts are studied in previous video semantic segmentation. However, there is no research about how to simultaneously learn static and motional contexts which are highly correlated and complementary to each other. To address this problem, we propose a Coarse-to-Fine Feature Mining (CFFM) technique to learn a unified presentation of static contexts and motional contexts. This technique consists of two parts: coarse-to-fine feature assembling and cross-frame feature mining. The former operation prepares data for further processing, enabling the subsequent joint learning of static and motional contexts. The latter operation mines useful information/contexts from the sequential frames to enhance the video contexts of the features of the target frame. The enhanced features can be directly applied for the final prediction. Experimental results on popular benchmarks demonstrate that the proposed CFFM performs favorably against state-of-the-art methods for video semantic segmentation. Our implementation is available at https://github.com/GuoleiSun/VSS-CFFM.

Adapting to a continuously evolving environment is a safety-critical challenge inevitably faced by all autonomous-driving systems. Existing image- and video-based driving datasets, however, fall short of capturing the mutable nature of the real world. In this paper, we introduce the largest multi-task synthetic dataset for autonomous driving, SHIFT. It presents discrete and continuous shifts in cloudiness, rain and fog intensity, time of day, and vehicle and pedestrian density. Featuring a comprehensive sensor suite and annotations for several mainstream perception tasks, SHIFT allows to investigate how a perception systems' performance degrades at increasing levels of domain shift, fostering the development of continuous adaptation strategies to mitigate this problem and assessing the robustness and generality of a model. Our dataset and benchmark toolkit are publicly available at www.vis.xyz/shift.

We propose Probabilistic Warp Consistency, a weakly-supervised learning objective for semantic matching. Our approach directly supervises the dense matching scores predicted by the network, encoded as a conditional probability distribution. We first construct an image triplet by applying a known warp to one of the images in a pair depicting different instances of the same object class. Our probabilistic learning objectives are then derived using the constraints arising from the resulting image triplet. We further account for occlusion and background clutter present in real image pairs by extending our probabilistic output space with a learnable unmatched state. To supervise it, we design an objective between image pairs depicting different object classes. We validate our method by applying it to four recent semantic matching architectures. Our weakly-supervised approach sets a new state-of-the-art on four challenging semantic matching benchmarks. Lastly, we demonstrate that our objective also brings substantial improvements in the strongly-supervised regime, when combined with keypoint annotations.

Densely annotating LiDAR point clouds remains too expensive and time-consuming to keep up with the ever growing volume of data. While current literature focuses on fully-supervised performance, developing efficient methods that take advantage of realistic weak supervision have yet to be explored. In this paper, we propose using scribbles to annotate LiDAR point clouds and release ScribbleKITTI, the first scribble-annotated dataset for LiDAR semantic segmentation. Furthermore, we present a pipeline to reduce the performance gap that arises when using such weak annotations. Our pipeline comprises of three stand-alone contributions that can be combined with any LiDAR semantic segmentation model to achieve up to 95.7% of the fully-supervised performance while using only 8% labeled points. Our scribble annotations and code are available at github.com/ouenal/scribblekitti.

Different self-supervised tasks (SSL) reveal different features from the data. The learned feature representations can exhibit different performance for each downstream task. In this light, this work aims to combine Multiple SSL tasks (Multi-SSL) that generalizes wellfor all downstream tasks. For this study, we investigate binaural sounds and image data. For binaural sounds, we propose three SSL tasks namely, spatial alignment, temporal synchronization offore-ground objects and binaural sounds and temporal gap prediction. We investigate several approaches of Multi-SSL and give insights into the downstream task performance on video retrieval, spatial sound super resolution, and semantic prediction using OmniAudio dataset. Our experiments on binaural sound representations demonstrate that Multi-SSL via incremental learning (IL) of SSL tasks outperforms single SSL task models and fully supervised models in the downstream task performance. As a check of applicability on other modalities, we also formulate our Multi-SSL models for image representation learning and we use the recently proposed SSL tasks, MoCov2 and DenseCL. Here, Multi-SSL surpasses recent methods such as MoCov2, DenseCL and DetCo by 2.06%,3.27% and 1.19% on VOC07 classification and +2.83, +1.56 and +1.61 AP on COCO detection.

Since the rise of vision-language navigation (VLN), great progress has been made in instruction following - building a follower to navigate environments under the guidance of instructions. However, far less attention has been paid to the inverse task: instruction generation - learning a speaker to generate grounded descriptions for navigation routes. Existing VLN methods train a speaker independently and often treat it as a data augmentation tool to strengthen the follower, while ignoring rich cross-task relations. Here we describe an approach that learns the two tasks simultaneously and exploits their intrinsic correlations to boost the training of each: the follower judges whether the speaker-created instruction explains the original navigation route correctly, and vice versa. Without the need of aligned instruction-path pairs, such cycle-consistent learning scheme is complementary to task-specific training targets defined on labeled data, and can also be applied over unlabeled paths (sampled without paired instructions). Another agent, called creator is added to generate counterfactual environments. It greatly changes current scenes yet leaves novel items - which are vital for the execution of original instructions - unchanged. Thus more informative training scenes are synthesized and the three agents compose a powerful VLN learning system. Extensive experiments on a standard benchmark show that our approach improves the performance of various follower models and produces accurate navigation instructions.

To achieve disentangled image manipulation, previous works depend heavily on manual annotation. Meanwhile, the available manipulations are limited to a pre-defined set the models were trainedfor. We propose a novelframework, i.e., Predict, Prevent, and Evaluate (PPE), for disentangled text-driven image manipulation that requires little manual annotation while being applicable to a wide variety of ma-nipulations. Our method approaches the targets by deeply exploiting the power of the large-scale pre-trained vision-language model CLIP [32]. Concretely, we firstly Predict the possibly entangled attributes for a given text command. Then, based on the predicted attributes, we introduce an entanglement loss to Prevent entanglements during training. Finally, we propose a new evaluation metric to Evaluate the disentangled image manipulation. We verify the effectiveness of our method on the challenging face editing task. Extensive experiments show that the proposed PPE frame-work achieves much better quantitative and qualitative re-sults than the up-to-date StyleCLIP [31] baseline. Code is available at https://github.com/zipengxuc/PPE.

Multi-Object Tracking (MOT) is most often approached in the tracking-by-detection paradigm, where object detections are associated through time. The association step naturally leads to discrete optimization problems. As these optimization problems are often NP-hard, they can only be solved exactly for small instances on current hardware. Adiabatic quantum computing (AQC) offers a solution for this, as it has the potential to provide a considerable speedup on a range of NP-hard optimization problems in the near future. However, current MOT formulations are unsuitable for quantum computing due to their scaling properties. In this work, we therefore propose the first MOT formulation designed to be solved with AQC. We employ an Ising model that represents the quantum mechanical system implemented on the AQC. We show that our approach is competitive compared with state-of-the-art optimization-based approaches, even when using of-the-shelf integer programming solvers. Finally, we demonstrate that our MOT problem is already solvable on the current generation of real quantum computers for small examples, and analyze the properties of the measured solutions.

Prevalent semantic segmentation solutions, despite their different network designs (FCN based or attention based) and mask decoding strategies (parametric softmax based or pixel-query based), can be placed in one category, by considering the softmax weights or query vectors as learnable class prototypes. In light of this prototype view, this study uncovers several limitations of such parametric segmentation regime, and proposes a nonparametric alternative based on non-learnable prototypes. Instead of prior methods learning a single weight/query vector for each class in a fully parametric manner, our model represents each class as a set of non-learnable prototypes, relying solely on the mean fea-tures of several training pixels within that class. The dense prediction is thus achieved by nonparametric nearest prototype retrieving. This allows our model to directly shape the pixel embedding space, by optimizing the arrangement between embedded pixels and anchored prototypes. It is able to handle arbitrary number of classes with a constant amount of learnable parameters. We empirically show that, with FCN based and attention based segmentation models (i.e., HR-Net, Swin, SegFormer) and backbones (i.e., ResNet, HRNet, Swin, MiT), our nonparametric framework yields compel-ling results over several datasets (i.e., ADE20K, Cityscapes, COCO-Stuff), and performs well in the large-vocabulary situation. We expect this work will provoke a rethink of the current de facto semantic segmentation model design.

Unsupervised domain adaptation (UDA) aims to adapt a model trained on the source domain (e.g. synthetic data) to the target domain (e.g. real-world data) without requiring further annotations on the target domain. This work focuses on UDA for semantic segmentation as real-world pixel-wise annotations are particularly expensive to acquire. As UDA methods for semantic segmentation are usually GPU memory intensive, most previous methods operate only on downscaled images. We question this design as low-resolution predictions often fail to preserve fine details. The alternative of training with random crops of high-resolution images alleviates this problem but falls short in capturing long-range, domain-robust context information. Therefore, we propose HRDA, a multi-resolution training approach for UDA, that combines the strengths of small high-resolution crops to preserve fine segmentation details and large low-resolution crops to capture long-range context dependencies with a learned scale attention, while maintaining a manageable GPU memory footprint. HRDA enables adapting small objects and preserving fine segmentation details. It significantly improves the state-of-the-art performance by 5.5 mIoU for GTA $$\rightarrow $$ Cityscapes and 4.9 mIoU for Synthia $$\rightarrow $$ Cityscapes, resulting in unprecedented 73.8 and 65.8 mIoU, respectively. The implementation is available at github.com/lhoyer/HRDA .

Exploiting similar and sharper scene patches in spatio-temporal neighborhoods is critical for video deblurring. However, CNN-based methods show limitations in capturing long-range dependencies and modeling non-local self-similarity. In this paper, we propose a novel framework, Flow-Guided Sparse Transformer (FGST), for video deblurring. In FGST, we customize a self-attention module, Flow-Guided Sparse Window-based Multi-head Self-Attention (FGSW-MSA). For each $query$ element on the blurry reference frame, FGSW-MSA enjoys the guidance of the estimated optical flow to globally sample spatially sparse yet highly related $key$ elements corresponding to the same scene patch in neighboring frames. Besides, we present a Recurrent Embedding (RE) mechanism to transfer information from past frames and strengthen long-range temporal dependencies. Comprehensive experiments demonstrate that our proposed FGST outperforms state-of-the-art (SOTA) methods on both DVD and GOPRO datasets and yields visually pleasant results in real video deblurring. https://github.com/linjing7/VR-Baseline

How to properly model the inter-frame relation within the video sequence is an important but unsolved challenge for video restoration (VR). In this work, we propose an unsupervised flow-aligned sequence-to-sequence model (S2SVR) to address this problem. On the one hand, the sequence-to-sequence model, which has proven capable of sequence modeling in the field of natural language processing, is explored for the first time in VR. Optimized serialization modeling shows potential in capturing long-range dependencies among frames. On the other hand, we equip the sequence-to-sequence model with an unsupervised optical flow estimator to maximize its potential. The flow estimator is trained with our proposed unsupervised distillation loss, which can alleviate the data discrepancy and inaccurate degraded optical flow issues of previous flow-based methods. With reliable optical flow, we can establish accurate correspondence among multiple frames, narrowing the domain difference between 1D language and 2D misaligned frames and improving the potential of the sequence-to-sequence model. S2SVR shows superior performance in multiple VR tasks, including video deblurring, video super-resolution, and compressed video quality enhancement. https://github.com/linjing7/VR-Baseline

A set of novel approaches for estimating epistemic uncertainty in deep neural networks with a single forward pass has recently emerged as a valid alternative to Bayesian Neural Networks. On the premise of informative representations, these deterministic uncertainty methods (DUMs) achieve strong performance on detecting out-of-distribution (OOD) data while adding negligible computational costs at inference time. However, it remains unclear whether DUMs are well calibrated and can seamlessly scale to real-world applications - both prerequisites for their practical deployment. To this end, we first provide a taxonomy of DUMs, and evaluate their calibration under continuous distributional shifts. Then, we extend them to semantic segmentation. We find that, while DUMs scale to realistic vision tasks and perform well on OOD detection, the practicality of current methods is undermined by poor calibration under distributional shifts.

The increased importance of mobile photography created a need for fast and performant RAW image processing pipelines capable of producing good visual results in spite of the mobile camera sensor limitations. While deep learning-based approaches can efficiently solve this problem, their computational requirements usually remain too large for high-resolution on-device image processing. To address this limitation, we propose a novel PyNET-V2 Mobile CNN architecture designed specifically for edge devices, being able to process RAW 12MP photos directly on mobile phones under 1.5 second and producing high perceptual photo quality. To train and to evaluate the performance of the proposed solution, we use the real-world Fujifilm UltraISP dataset consisting on thousands of RAW-RGB image pairs captured with a professional medium-format 102MP Fujifilm camera and a popular Sony mobile camera sensor. The results demonstrate that the PyNET-V2 Mobile model can substantially surpass the quality of tradition ISP pipelines, while outperforming the previously introduced neural network-based solutions designed for fast image processing. Furthermore, we show that the proposed architecture is also compatible with the latest mobile AI accelerators such as NPUs or APUs that can be used to further reduce the latency of the model to as little as 0.5 second. The dataset, code and pre-trained models used in this paper are available on the project website: https://github.com/gmalivenko/PyNET-v2

This paper proposes a Perceptual Learned Video Compression (PLVC) approach with recurrent conditional GAN. We employ the recurrent auto-encoder-based compression network as the generator, and most importantly, we propose a recurrent conditional discriminator, which judges raw vs. compressed video conditioned on both spatial and temporal features, including the latent representation, temporal motion and hidden states in recurrent cells. This way, the adversarial training pushes the generated video to be not only spatially photo-realistic but also temporally consistent with the groundtruth and coherent among video frames. The experimental results show that the learned PLVC model compresses video with good perceptual quality at low bit-rate, and that it outperforms the official HEVC test model (HM 16.20) and the existing learned video compression approaches for several perceptual quality metrics and user studies. The project page is available at https://github.com/RenYang-home/PLVC.

In coded aperture snapshot spectral compressive imaging (CASSI) systems, hyperspectral image (HSI) reconstruction methods are employed to recover the spatial-spectral signal from a compressed measurement. Among these algorithms, deep unfolding methods demonstrate promising performance but suffer from two issues. Firstly, they do not estimate the degradation patterns and ill-posedness degree from CASSI to guide the iterative learning. Secondly, they are mainly CNN-based, showing limitations in capturing long-range dependencies. In this paper, we propose a principled Degradation-Aware Unfolding Framework (DAUF) that estimates parameters from the compressed image and physical mask, and then uses these parameters to control each iteration. Moreover, we customize a novel Half-Shuffle Transformer (HST) that simultaneously captures local contents and non-local dependencies. By plugging HST into DAUF, we establish the first Transformer-based deep unfolding method, Degradation-Aware Unfolding Half-Shuffle Transformer (DAUHST), for HSI reconstruction. Experiments show that DAUHST surpasses state-of-the-art methods while requiring cheaper computational and memory costs. Code and models are publicly available at https://github.com/caiyuanhao1998/MST

Video restoration aims at restoring multiple high-quality frames from multiple low-quality frames. Existing video restoration methods generally fall into two extreme cases, i.e., they either restore all frames in parallel or restore the video frame by frame in a recurrent way, which would result in different merits and drawbacks. Typically, the former has the advantage of temporal information fusion. However, it suffers from large model size and intensive memory consumption; the latter has a relatively small model size as it shares parameters across frames; however, it lacks long-range dependency modeling ability and parallelizability. In this paper, we attempt to integrate the advantages of the two cases by proposing a recurrent video restoration transformer, namely RVRT. RVRT processes local neighboring frames in parallel within a globally recurrent framework which can achieve a good trade-off between model size, effectiveness, and efficiency. Specifically, RVRT divides the video into multiple clips and uses the previously inferred clip feature to estimate the subsequent clip feature. Within each clip, different frame features are jointly updated with implicit feature aggregation. Across different clips, the guided deformable attention is designed for clip-to-clip alignment, which predicts multiple relevant locations from the whole inferred clip and aggregates their features by the attention mechanism. Extensive experiments on video super-resolution, deblurring, and denoising show that the proposed RVRT achieves state-of-the-art performance on benchmark datasets with balanced model size, testing memory and runtime.

Despite the tremendous progress in zero-shot learning (ZSL), the majority of existing methods still rely on human-annotated attributes, which are difficult to annotate and scale. An unsupervised alternative is to represent each class using the word embedding associated with its semantic class name. However, word embeddings extracted from pre-trained language models do not necessarily capture visual similarities, resulting in poor zero-shot performance. In this work, we argue that online textual documents e.g., Wikipedia, contain rich visual descriptions about object classes, therefore can be used as powerful unsupervised side information for ZSL. To this end, we propose I2DFormer, a novel transformer-based ZSL framework that jointly learns to encode images and documents by aligning both modalities in a shared embedding space. In order to distill discriminative visual words from noisy documents, we introduce a new cross-modal attention module that learns fine-grained interactions between image patches and document words. Consequently, our I2DFormer not only learns highly discriminative document embeddings that capture visual similarities but also gains the ability to localize visually relevant words in image regions. Quantitatively, we demonstrate that our I2DFormer significantly outperforms previous unsupervised semantic embeddings under both zero-shot and generalized zero-shot learning settings on three public datasets. Qualitatively, we show that our method leads to highly interpretable results where document words can be grounded in the image regions.

With the emergence of varied visual navigation tasks (e.g., image-/object-/audio-goal and vision-language navigation) that specify the target in different ways, the community has made appealing advances in training specialized agents capable of handling individual navigation tasks well. Given plenty of embodied navigation tasks and task-specific solutions, we address a more fundamental question: can we learn a single powerful agent that masters not one but multiple navigation tasks concurrently? First, we propose VXN, a large-scale 3D dataset that instantiates~four classic navigation tasks in standardized, continuous, and audiovisual-rich environments. Second, we propose Vienna, a versatile embodied navigation agent that simultaneously learns to perform the four navigation tasks with one model. Building upon a full-attentive architecture, Vienna formulates various navigation tasks as a unified, parse-and-query procedure: the target description, augmented with four task embeddings, is comprehensively interpreted into a set of diversified goal vectors, which are refined as the navigation progresses, and used as queries to retrieve supportive context from episodic history for decision making. This enables the reuse of knowledge across navigation tasks with varying input domains/modalities. We empirically demonstrate that, compared with learning each visual navigation task individually, our multitask agent achieves comparable or even better performance with reduced complexity.

We study the problem of robust domain adaptation in the context of unavailable target labels and source data. The considered robustness is against adversarial perturbations. This paper aims at answering the question of finding the right strategy to make the target model robust and accurate in the setting of unsupervised domain adaptation without source data. The major findings of this paper are: (i) robust source models can be transferred robustly to the target; (ii) robust domain adaptation can greatly benefit from nonrobust pseudo-labels and the pair-wise contrastive loss. The proposed method of using non-robust pseudo-labels performs surprisingly well on both clean and adversarial samples, for the task of image classification. We show a consistent performance improvement of over 10% in accuracy against the tested baselines on four benchmark datasets. Our source code will be made publicly available.

We present a modern solution to the multi-view photometric stereo problem (MVPS). Our work suitably exploits the image formation model in a MVPS experimental setup to recover the dense 3D reconstruction of an object from images. We procure the surface orientation using a photometric stereo (PS) image formation model and blend it with a multi-view neural radiance field representation to recover the object’s surface geometry. Contrary to the previous multi-staged framework to MVPS, where the position, iso-depth contours, or orientation measurements are estimated independently and then fused later, our method is simple to implement and realize. Our method performs neural rendering of multi-view images while utilizing surface normals estimated by a deep photometric stereo network. We render the MVPS images by considering the object’s surface normals for each 3D sample point along the viewing direction rather than explicitly using the density gradient in the volume space via 3D occupancy information. We optimize the proposed neural radiance field representation for the MVPS setup efficiently using a fully connected deep network to recover the 3D geometry of an object. Extensive evaluation on the DiLiGenT-MV benchmark dataset shows that our method performs better than the approaches that perform only PS or only multi-view stereo (MVS) and provides comparable results against the state-of-the-art multistage fusion methods.

This work studies Hyperspectral image (HSI) super-resolution (SR). HSI SR is characterized by high-dimensional data and a limited amount of training examples. This raises challenges for training deep neural networks that are known to be data hungry. This work addresses this issue with two contributions. First, we observe that HSI SR and RGB image SR are correlated and develop a novel multi-tasking network to train them jointly so that the auxiliary task RGB image SR can provide additional supervision and regulate the network training. Second, we extend the network to a semi-supervised setting so that it can learn from datasets containing only low-resolution HSIs. With these contributions, our method is able to learn hyperspectral image super-resolution from heterogeneous datasets and lifts the requirement for having a large amount of high resolution (HR) HSI training samples. Extensive experiments on three standard datasets show that our method outperforms existing methods significantly and underpin the relevance of our contributions. Our code can be found at https://github.com/kli8996/HSISR.git.

Super-resolution is an ill-posed problem, where a ground-truth high-resolution image represents only one possibility in the space of plausible solutions. Yet, the dominant paradigm is to employ pixel-wise losses, such as L1, which drive the prediction towards a blurry average. This leads to fundamentally conflicting objectives when combined with adversarial losses, which degrades the final quality. We address this issue by revisiting the L1loss and show that it corresponds to a one-layer conditional flow. Inspired by this relation, we explore general flows as a fidelity-based alternative to the L1objective. We demonstrate that the flexibility of deeper flows leads to better visual quality and consistency when combined with adversarial losses. We conduct extensive user studies for three datasets and scale factors, where our approach is shown to outperform state-of-the-art methods for photo-realistic super-resolution. Code and trained models: git.io/AdFlow

We present an automated machine learning approach for uncalibrated photometric stereo (PS). Our work aims at discovering lightweight and computationally efficient PS neural networks with excellent surface normal accuracy. Unlike previous uncalibrated deep PS networks, which are handcrafted and carefully tuned, we leverage differentiable neural architecture search (NAS) strategy to find uncalibrated PS architecture automatically. We begin by defining a discrete search space for a light calibration network and a normal estimation network, respectively. We then perform a continuous relaxation of this search space and present a gradient-based optimization strategy to find an efficient light calibration and normal estimation network. Directly applying the NAS methodology to uncalibrated PS is not straightforward as certain task-specific constraints must be satisfied, which we impose explicitly. Moreover, we search for and train the two networks separately to account for the Generalized Bas-Relief (GBR) ambiguity. Extensive experiments on the DiLiGenT dataset show that the automatically searched neural architectures performance compares favorably with the state-of-the-art uncalibrated PS methods while having a lower memory footprint.

Generic object counting in natural scenes is a challenging computer vision problem. Existing approaches either rely on instance-level supervision or absolute count information to train a generic object counter. We introduce a partially supervised setting that significantly reduces the supervision level required for generic object counting. We propose two novel frameworks, named lower-count (LC) and reduced lower-count (RLC), to enable object counting under this setting. Our frameworks are built on a novel dual-branch architecture that has an image classification and a density branch. Our LC framework reduces the annotation cost due to multiple instances in an image by using only lower-count supervision for all object categories. Our RLC framework further reduces the annotation cost arising from large numbers of object categories in a dataset by only using lower-count supervision for a subset of categories and class-labels for the remaining ones. The RLC framework extends our dual-branch LC framework with a novel weight modulation layer and a category-independent density map prediction. Experiments are performed on COCO, Visual Genome and PASCAL 2007 datasets. Our frameworks perform on par with state-of-the-art approaches using higher levels of supervision. Additionally, we demonstrate the applicability of our LC supervised density map for image-level supervised instance segmentation.

In this article, we model a set of pixelwise object segmentation tasks - automatic video segmentation (AVS), image co-segmentation (ICS) and few-shot semantic segmentation (FSS) - in a unified view of segmenting objects from relational visual data. To this end, we propose an attentive graph neural network (AGNN) that addresses these tasks in a holistic fashion, by formulating them as a process of iterative information fusion over data graphs. It builds a fully-connected graph to efficiently represent visual data as nodes and relations between data instances as edges. The underlying relations are described by a differentiable attention mechanism, which thoroughly examines fine-grained semantic similarities between all the possible location pairs in two data instances. Through parametric message passing, AGNN is able to capture knowledge from the relational visual data, enabling more accurate object discovery and segmentation. Experiments show that AGNN can automatically highlight primary foreground objects from video sequences (i.e., automatic video segmentation), and extract common objects from noisy collections of semantically related images (i.e., image co-segmentation). AGNN can even generalize segment new categories with little annotated data (i.e., few-shot semantic segmentation). Taken together, our results demonstrate that AGNN provides a powerful tool that is applicable to a wide range of pixel-wise object pattern understanding tasks with relational visual data. Our algorithm implementations have been made publicly available at https://github.com/carrierlxk/AGNN.

It is quite laborious and costly to manually label LiDAR point cloud data for training high-quality 3D object detectors. This work proposes a weakly supervised framework which allows learning 3D detection from a few weakly annotated examples. This is achieved by a two-stage architecture design. Stage-1 learns to generate cylindrical object proposals under inaccurate and inexact supervision, obtained by our proposed BEV center-click annotation strategy, where only the horizontal object centers are click-annotated in bird's view scenes. Stage-2 learns to predict cuboids and confidence scores in a coarse-to-fine, cascade manner, under incomplete supervision, i.e., only a small portion of object cuboids are precisely annotated. With KITTI dataset, using only 500 weakly annotated scenes and 534 precisely labeled vehicle instances, our method achieves 86-97 percent the performance of current top-leading, fully supervised detectors (which require 3,712 exhaustively annotated scenes with 15,654 instances). More importantly, with our elaborately designed network architecture, our trained model can be applied as a 3D object annotator, supporting both automatic and active (human-in-the-loop) working modes. The annotations generated by our model can be used to train 3D object detectors, achieving over 95 percent of their original performance (with manually labeled training data). Our experiments also show our model's potential in boosting performance when given more training data. The above designs make our approach highly practical and open-up opportunities for learning 3D detection at reduced annotation cost.

We address the problem of semantic nighttime image segmentation and improve the state-of-the-art, by adapting daytime models to nighttime without using nighttime annotations. Moreover, we design a new evaluation framework to address the substantial uncertainty of semantics in nighttime images. Our central contributions are: 1) a curriculum framework to gradually adapt semantic segmentation models from day to night through progressively darker times of day, exploiting cross-time-of-day correspondences between daytime images from a reference map and dark images to guide the label inference in the dark domains; 2) a novel uncertainty-aware annotation and evaluation framework and metric for semantic segmentation, including image regions beyond human recognition capability in the evaluation in a principled fashion; 3) the Dark Zurich dataset, comprising 2416 unlabeled nighttime and 2920 unlabeled twilight images with correspondences to their daytime counterparts plus a set of 201 nighttime images with fine pixel-level annotations created with our protocol, which serves as a first benchmark for our novel evaluation. Experiments show that our map-guided curriculum adaptation significantly outperforms state-of-the-art methods on nighttime sets both for standard metrics and our uncertainty-aware metric. Furthermore, our uncertainty-aware evaluation reveals that selective invalidation of predictions can improve results on data with ambiguous content such as our benchmark and profit safety-oriented applications involving invalid inputs.

With the advent of deep learning, many dense prediction tasks, i.e., tasks that produce pixel-level predictions, have seen significant performance improvements. The typical approach is to learn these tasks in isolation, that is, a separate neural network is trained for each individual task. Yet, recent multi-task learning (MTL) techniques have shown promising results w.r.t. performance, computations and/or memory footprint, by jointly tackling multiple tasks through a learned shared representation. In this survey, we provide a well-rounded view on state-of-the-art deep learning approaches for MTL in computer vision, explicitly emphasizing on dense prediction tasks. Our contributions concern the following. First, we consider MTL from a network architecture point-of-view. We include an extensive overview and discuss the advantages/disadvantages of recent popular MTL models. Second, we examine various optimization methods to tackle the joint learning of multiple tasks. We summarize the qualitative elements of these works and explore their commonalities and differences. Finally, we provide an extensive experimental evaluation across a variety of dense prediction benchmarks to examine the pros and cons of the different methods, including both architectural and optimization based strategies.

Recent works on plug-and-play image restoration have shown that a denoiser can implicitly serve as the image prior for model-based methods to solve many inverse problems. Such a property induces considerable advantages for plug-and-play image restoration (e.g., integrating the flexibility of model-based method and effectiveness of learning-based methods) when the denoiser is discriminatively learned via deep convolutional neural network (CNN) with large modeling capacity. However, while deeper and larger CNN models are rapidly gaining popularity, existing plug-and-play image restoration hinders its performance due to the lack of suitable denoiser prior. In order to push the limits of plug-and-play image restoration, we set up a benchmark deep denoiser prior by training a highly flexible and effective CNN denoiser. We then plug the deep denoiser prior as a modular part into a half quadratic splitting based iterative algorithm to solve various image restoration problems. We, meanwhile, provide a thorough analysis of parameter setting, intermediate results and empirical convergence to better understand the working mechanism. Experimental results on three representative image restoration tasks, including deblurring, super-resolution and demosaicing, demonstrate that the proposed plug-and-play image restoration with deep denoiser prior not only significantly outperforms other state-of-the-art model-based methods but also achieves competitive or even superior performance against state-of-the-art learning-based methods. The source code is available at https://github.com/cszn/DPIR.

Dense human pose estimation is the problem of learning dense correspondences between RGB images and the surfaces of human bodies, which finds various applications, such as human body reconstruction, human pose transfer, and human action recognition. Prior dense pose estimation methods are all based on Mask R-CNN framework and operate in a top-down manner of first attempting to identify a bounding box for each person and matching dense correspondences in each bounding box. Consequently, these methods lack robustness due to their critical dependence on the Mask R-CNN detection, and the runtime increases drastically as the number of persons in the image increases. We therefore propose a novel alternative method for solving the dense pose estimation problem, called Direct Dense Pose (DDP). DDP first predicts the instance mask and global IUV representation separately and then combines them together. We also propose a simple yet effective 2D temporal-smoothing scheme to alleviate the temporal jitters when dealing with video data. Experiments demonstrate that DDP overcomes the limitations of previous top-down baseline methods and achieves competitive accuracy. In addition, DDP is computationally more efficient than previous dense pose estimation methods, and it reduces jitters when applied to a video sequence, which is a problem plaguing the previous methods.

Recently Normalizing Flows (NFs) have demonstrated state-of-the-art performance on modeling 3D point clouds while allowing sampling with arbitrary resolution at inference time. However, these flow-based models still have fundamental limitations on complicated geometries. This work generalizes prior work by introducing additional discrete latent variable, i.e. mixture model. This circumvents limitations of prior approaches, leads to more parameter efficient models and reduces the inference runtime. Moreover, in this more general framework each component learns to specialize in a particular subregion of an object in a completely unsupervised fashion yielding promising clustering properties. We further demonstrate that by adding data augmentation, individual mixture components can learn to specialize in a semantically meaningful manner. We evaluate mixtures of NFs on generation, autoencoding and single-view reconstruction based on the ShapeNet dataset.

Monocular depth reconstruction of complex and dynamic scenes is a highly challenging problem. While for rigid scenes learning-based methods have been offering promising results even in unsupervised cases, there exists little to no literature addressing the same for dynamic and deformable scenes. In this work, we present an unsupervised monocular framework for dense depth estimation of dynamic scenes, which jointly reconstructs rigid and nonrigid parts without explicitly modelling the camera motion. Using dense correspondences, we derive a training objective that aims to opportunistically preserve pairwise distances between reconstructed 3D points. In this process, the dense depth map is learned implicitly using the as-rigid-as-possible hypothesis. Our method provides promising results, demonstrating its capability of reconstructing 3D from challenging videos of non-rigid scenes. Furthermore, the proposed method also provides unsupervised motion segmentation results as an auxiliary output.

Image-to-image translation is to map images from a given style to another given style. While exceptionally successful, current methods assume the availability of training images in both source and target domains, which does not always hold in practice. Inspired by humans' reasoning capability of analogy, we propose analogical image translation (AIT) that exploit the concept of gist, for the first time. Given images of two styles in the source domain: A and A', along with images B of the first style in the target domain, learn a model to translate B to B' in the target domain, such that A:A' :: B:B'. AIT is especially useful for translation scenarios in which training data of one style is hard to obtain but training data of the same two styles in another domain is available. For instance, in the case from normal conditions to extreme, rare conditions, obtaining real training images for the latter case is challenging. However, obtaining synthetic data for both cases is relatively easy. In this work, we aim at adding adverse weather effects, more specifically fog, to images taken in clear weather. To circumvent the challenge of collecting real foggy images, AIT learns the gist of translating synthetic clear-weather to foggy images, followed by adding fog effects onto real clear-weather images, without ever seeing any real foggy image. AIT achieves zero-shot image translation capability, whose effectiveness and benefit are demonstrated by the downstream task of semantic foggy scene understanding.

This paper aims at enlarging the problem of Neural Architecture Search (NAS) from Single-Path and Multi-Path Search to automated Mixed-Path Search. In particular, we model the NAS problem as a sparse supernet using a new continuous architecture representation with a mixture of sparsity constraints. The sparse supernet enables us to automatically achieve sparsely-mixed paths upon a compact set of nodes. To optimize the proposed sparse supernet, we exploit a hierarchical accelerated proximal gradient algorithm within a bi-level optimization framework. Extensive experiments on Convolutional Neural Network and Recurrent Neural Network search demonstrate that the proposed method is capable of searching for compact, general and powerful neural architectures.

We study low-rank parameterizations of weight matrices with embedded spectral properties in the Deep Learning context. The low-rank property leads to parameter efficiency and permits taking computational shortcuts when computing mappings. Spectral properties are often subject to constraints in optimization problems, leading to better models and stability of optimization. We start by looking at the compact SVD parameterization of weight matrices and identifying redundancy sources in the parameterization. We further apply the Tensor Train (TT) decomposition to the compact SVD components, and propose a non-redundant differentiable parameterization of fixed TT-rank tensor manifolds, termed the Spectral Tensor Train Parameterization (STTP). We demonstrate the effects of neural network compression in the image classification setting, and both compression and improved training stability in the generative adversarial training setting. Project website: www.obukhov.ai/sttp

While single-image super-resolution (SISR) has attracted substantial interest in recent years, the proposed approaches are limited to learning image priors in order to add high frequency details. In contrast, multi-frame super-resolution (MFSR) offers the possibility of reconstructing rich details by combining signal information from multiple shifted images. This key advantage, along with the increasing popularity of burst photography, have made MFSR an important problem for real-world applications.We propose a novel architecture for the burst super-resolution task. Our network takes multiple noisy RAW images as input, and generates a denoised, super-resolved RGB image as output. This is achieved by explicitly aligning deep embeddings of the input frames using pixel-wise optical flow. The information from all frames are then adaptively merged using an attention-based fusion module. In order to enable training and evaluation on real-world data, we additionally introduce the BurstSR dataset, consisting of smartphone bursts and high-resolution DSLR ground-truth. We perform comprehensive experimental analysis, demonstrating the effectiveness of the proposed architecture.

While state-of-the-art 3D Convolutional Neural Networks (CNN) achieve very good results on action recognition datasets, they are computationally very expensive and require many GFLOPs. While the GFLOPs of a 3D CNN can be decreased by reducing the temporal feature resolution within the network, there is no setting that is optimal for all input clips. In this work, we therefore introduce a differentiable Similarity Guided Sampling (SGS) module, which can be plugged into any existing 3D CNN architecture. SGS empowers 3D CNNs by learning the similarity of temporal features and grouping similar features together. As a result, the temporal feature resolution is not anymore static but it varies for each input video clip. By integrating SGS as an additional layer within current 3D CNNs, we can convert them into much more efficient 3D CNNs with adaptive temporal feature resolutions (ATFR). Our evaluations show that the proposed module improves the state-of-the-art by reducing the computational cost (GFLOPs) by half while preserving or even improving the accuracy. We evaluate our module by adding it to multiple state-of-the-art 3D CNNs on various datasets such as Kinetics-600, Kinetics-400, mini-Kinetics, Something-Something V2, UCF101, and HMDB51.

Open compound domain adaptation (OCDA) is a domain adaptation setting, where target domain is modeled as a compound of multiple unknown homogeneous domains, which brings the advantage of improved generalization to unseen domains. In this work, we propose a principled meta-learning based approach to OCDA for semantic segmentation, MOCDA, by modeling the unlabeled target domain continuously. Our approach consists of four key steps. First, we cluster target domain into multiple sub-target domains by image styles, extracted in an unsupervised manner. Then, different sub-target domains are split into independent branches, for which batch normalization parameters are learnt to treat them independently. A meta-learner is thereafter deployed to learn to fuse sub-target domain-specific predictions, conditioned upon the style code. Meanwhile, we learn to online update the model by model-agnostic meta-learning (MAML) algorithm, thus to further improve generalization. We validate the benefits of our approach by extensive experiments on synthetic-to-real knowledge transfer benchmark, where we achieve the state-of-the-art performance in both compound and open domains.

Training deep networks for semantic segmentation requires large amounts of labeled training data, which presents a major challenge in practice, as labeling segmentation masks is a highly labor-intensive process. To address this issue, we present a framework for semi-supervised semantic segmentation, which is enhanced by self-supervised monocular depth estimation from unlabeled image sequences. In particular, we propose three key contributions: (1) We transfer knowledge from features learned during self-supervised depth estimation to semantic segmentation, (2) we implement a strong data augmentation by blending images and labels using the geometry of the scene, and (3) we utilize the depth feature diversity as well as the level of difficulty of learning depth in a student-teacher framework to select the most useful samples to be annotated for semantic segmentation. We validate the proposed model on the Cityscapes dataset, where all three modules demonstrate significant performance gains, and we achieve state-of-the-art results for semi-supervised semantic segmentation. The implementation is available at https://github.com/lhoyer/improving_segmentation_with_selfsupervised_depth.

This paper presents an uncalibrated deep neural network framework for the photometric stereo problem. For training models to solve the problem, existing neural network-based methods either require exact light directions or ground-truth surface normals of the object or both. However, in practice, it is challenging to procure both of this information precisely, which restricts the broader adoption of photometric stereo algorithms for vision application. To bypass this difficulty, we propose an uncalibrated neural inverse rendering approach to this problem. Our method first estimates the light directions from the input images and then optimizes an image reconstruction loss to calculate the surface normals, bidirectional reflectance distribution function value, and depth. Additionally, our formulation explicitly models the concave and convex parts of a complex surface to consider the effects of interreflections in the image formation process. Extensive evaluation of the proposed method on the challenging subjects generally shows comparable or better results than the supervised and classical approaches.

In this paper, we tackle the problem of convolutional neural network design. Instead of focusing on the design of the overall architecture, we investigate a design space that is usually overlooked, i.e. adjusting the channel configurations of predefined networks. We find that this adjustment can be achieved by shrinking widened baseline networks and leads to superior performance. Based on that, we articulate the "heterogeneity hypothesis": with the same training protocol, there exists a layer-wise differentiated net-work architecture (LW-DNA) that can outperform the original network with regular channel configurations but with a lower level of model complexity.The LW-DNA models are identified without extra computational cost or training time compared with the original network. This constraint leads to controlled experiments which direct the focus to the importance of layer-wise specific channel configurations. LW-DNA models come with advantages related to overfitting, i.e. the relative relationship between model complexity and dataset size. Experiments are conducted on various networks and datasets for image classification, visual tracking and image restoration. The resultant LW-DNA models consistently outperform the baseline models. Code is available at https://github.com/ofsoundof/Heterogeneity_Hypothesis.git.

Kernel estimation is generally one of the key problems for blind image super-resolution (SR). Recently, Double-DIP proposes to model the kernel via a network architecture prior, while KernelGAN employs the deep linear network and several regularization losses to constrain the kernel space. However, they fail to fully exploit the general SR kernel assumption that anisotropic Gaussian kernels are sufficient for image SR. To address this issue, this paper proposes a normalizing flow-based kernel prior (FKP) for kernel modeling. By learning an invertible mapping between the anisotropic Gaussian kernel distribution and a tractable latent distribution, FKP can be easily used to replace the kernel modeling modules of Double-DIP and KernelGAN. Specifically, FKP optimizes the kernel in the latent space rather than the network parameter space, which allows it to generate reasonable kernel initialization, traverse the learned kernel manifold and improve the optimization stability. Extensive experiments on synthetic and real-world images demonstrate that the proposed FKP can significantly improve the kernel estimation accuracy with less parameters, runtime and memory usage, leading to state-of-the-art blind SR results.

We define the concept of CompositeTasking as the fusion of multiple, spatially distributed tasks, for various aspects of image understanding. Learning to perform spatially distributed tasks is motivated by the frequent availability of only sparse labels across tasks, and the desire for a compact multi-tasking network. To facilitate CompositeTasking, we introduce a novel task conditioning model – a single encoder-decoder network that performs multiple, spatially varying tasks at once. The proposed network takes an image and a set of pixel-wise dense task requests as inputs, and performs the requested prediction task for each pixel. Moreover, we also learn the composition of tasks that needs to be performed according to some CompositeTasking rules, which includes the decision of where to apply which task. It not only offers us a compact network for multitasking, but also allows for task-editing. Another strength of the proposed method is demonstrated by only having to supply sparse supervision per task. The obtained results are on par with our baselines that use dense supervision and a multi-headed multi-tasking design. The source code will be made publicly available at www.github.com/nikola3794/composite-tasking.

We present an approach for encoding visual task relationships to improve model performance in an Unsupervised Domain Adaptation (UDA) setting. Semantic segmentation and monocular depth estimation are shown to be complementary tasks; in a multi-task learning setting, a proper encoding of their relationships can further improve performance on both tasks. Motivated by this observation, we propose a novel Cross-Task Relation Layer (CTRL), which encodes task dependencies between the semantic and depth predictions. To capture the cross-task relationships, we propose a neural network architecture that contains task-specific and cross-task refinement heads. Furthermore, we propose an Iterative Self-Learning (ISL) training scheme, which exploits semantic pseudo-labels to provide extra supervision on the target domain. We experimentally observe improvements in both tasks’ performance because the complementary information present in these tasks is better captured. Specifically, we show that: (1) our approach improves performance on all tasks when they are complementary and mutually dependent; (2) the CTRL helps to improve both semantic segmentation and depth estimation tasks performance in the challenging UDA setting; (3) the proposed ISL training scheme further improves the semantic segmentation performance. The implementation is available at https://github.com/susaha/ctrl-uda.

Action segmentation refers to inferring boundaries of semantically consistent visual concepts in videos and is an important requirement for many video understanding tasks. For this and other video understanding tasks, supervised approaches have achieved encouraging performance but require a high volume of detailed frame-level annotations. We present a fully automatic and unsupervised approach for segmenting actions in a video that does not require any training. Our proposal is an effective temporally-weighted hierarchical clustering algorithm that can group semantically consistent frames of the video. Our main finding is that representing a video with a 1-nearest neighbor graph by taking into account the time progression is sufficient to form semantically and temporally consistent clusters of frames where each cluster may represent some action in the video. Additionally, we establish strong unsupervised baselines for action segmentation and show significant performance improvements over published unsupervised methods on five challenging action segmentation datasets. Our code is available.1

Generative adversarial networks (GANs) have shown impressive results in both unconditional and conditional image generation. In recent literature, it is shown that pre-trained GANs, on a different dataset, can be transferred to improve the image generation from a small target data. The same, however, has not been well-studied in the case of conditional GANs (cGANs), which provides new opportunities for knowledge transfer compared to unconditional setup. In particular, the new classes may borrow knowledge from the related old classes, or share knowledge among themselves to improve the training. This motivates us to study the problem of efficient conditional GAN transfer with knowledge propagation across classes. To address this problem, we introduce a new GAN transfer method to explicitly propagate the knowledge from the old classes to the new classes. The key idea is to enforce the popularly used conditional batch normalization (BN) to learn the class-specific information of the new classes from that of the old classes, with implicit knowledge sharing among the new ones. This allows for an efficient knowledge propagation from the old classes to the new ones, with the BN parameters increasing linearly with the number of new classes. The extensive evaluation demonstrates the clear superiority of the proposed method over state-of-the-art competitors for efficient conditional GAN transfer tasks. The code is available at: https://github.com/mshahbazi72/cGANTransfer

Establishing dense correspondences between a pair of images is an important and general problem. However, dense flow estimation is often inaccurate in the case of large displacements or homogeneous regions. For most applications and down-stream tasks, such as pose estimation, image manipulation, or 3D reconstruction, it is crucial to know when and where to trust the estimated matches.In this work, we aim to estimate a dense flow field relating two images, coupled with a robust pixel-wise confidence map indicating the reliability and accuracy of the prediction. We develop a flexible probabilistic approach that jointly learns the flow prediction and its uncertainty. In particular, we parametrize the predictive distribution as a constrained mixture model, ensuring better modelling of both accurate flow predictions and outliers. Moreover, we develop an architecture and training strategy tailored for robust and generalizable uncertainty prediction in the context of self-supervised training. Our approach obtains state- of-the-art results on multiple challenging geometric matching and optical flow datasets. We further validate the usefulness of our probabilistic confidence estimation for the task of pose estimation. Code and models are available at https://github.com/PruneTruong/PDCNet.

The difficulty of obtaining paired data remains a major bottleneck for learning image restoration and enhancement models for real-world applications. Current strategies aim to synthesize realistic training data by modeling noise and degradations that appear in real-world settings. We propose DeFlow, a method for learning stochastic image degradations from unpaired data. Our approach is based on a novel unpaired learning formulation for conditional normalizing flows. We model the degradation process in the latent space of a shared flow encoder-decoder network. This allows us to learn the conditional distribution of a noisy image given the clean input by solely minimizing the negative log-likelihood of the marginal distributions. We validate our DeFlow formulation on the task of joint image restoration and super-resolution. The models trained with the synthetic data generated by DeFlow outperform previous learnable approaches on three recent datasets. Code and trained models will be made available at: https://github.com/volflow/DeFlow

To address the challenging task of instance-aware human part parsing, a new bottom-up regime is proposed to learn category-level human semantic segmentation as well as multi-person pose estimation in a joint and end-to-end manner. It is a compact, efficient and powerful framework that exploits structural information over different human granularities and eases the difficulty of person partitioning. Specifically, a dense-to-sparse projection field, which allows explicitly associating dense human semantics with sparse keypoints, is learnt and progressively improved over the network feature pyramid for robustness. Then, the difficult pixel grouping problem is cast as an easier, multi-person joint assembling task. By formulating joint association as maximum-weight bipartite matching, a differentiable solution is developed to exploit projected gradient descent and Dykstra’s cyclic projection algorithm. This makes our method end-to-end trainable and allows back-propagating the grouping error to directly supervise multi-granularity human representation learning. This is distinguished from current bottom-up human parsers or pose estimators which require sophisticated post-processing or heuristic greedy algorithms. Experiments on three instance-aware human parsing datasets show that our model outperforms other bottom-up alternatives with much more efficient inference.

Humans can communicate emotions through a plethora of facial expressions, each with its own intensity, nuances and ambiguities. The generation of such variety by means of conditional GANs is limited to the expressions encoded in the used label system. These limitations are caused either due to burdensome labelling demand or the confounded label space. On the other hand, learning from inexpensive and intuitive basic categorical emotion labels leads to limited emotion variability. In this paper, we propose a novel GAN-based framework that learns an expressive and interpretable conditional space (usable as a label space) of emotions, instead of conditioning on handcrafted labels. Our framework only uses the categorical labels of basic emotions to learn jointly the conditional space as well as emotion manipulation. Such learning can benefit from the image variability within discrete labels, especially when the intrinsic labels reside beyond the discrete space of the defined. Our experiments demonstrate the effectiveness of the proposed framework, by allowing us to control and generate a gamut of complex and compound emotions while using only the basic categorical emotion labels during training. Our source code is available at https://github.com/stefanodapolito/GANmut.

We address the problem of face anti-spoofing which aims to make the face verification systems robust in the real world settings. The context of detecting live vs. spoofed face images may differ significantly in the target domain, when compared to that of labeled source domain where the model is trained. Such difference may be caused due to new and unknown spoof types, illumination conditions, scene backgrounds, among many others. These varieties of differences make the target a compound domain, thus calling for the problem of the unsupervised compound domain adaptation. We demonstrate the effectiveness of the compound domain assumption for the task of face anti-spoofing, for the first time in this work. To this end, we propose a memory augmentation method for adapting the source model to the target domain in a domain aware manner, inspired by [29]. The adaptation process is further improved by using the curriculum learning and the domain agnostic source network training approaches. The proposed method successfully adapts to the compound target domain consisting multiple new spoof types. Our experiments on multiple benchmark datasets demonstrate the superiority of the proposed method over the state-of-the-art. Our source code will be made publicly available.

The presence of objects that are confusingly similar to the tracked target, poses a fundamental challenge in appearance-based visual tracking. Such distractor objects are easily misclassified as the target itself, leading to eventual tracking failure. While most methods strive to suppress distractors through more powerful appearance models, we take an alternative approach.We propose to keep track of distractor objects in order to continue tracking the target. To this end, we introduce a learned association network, allowing us to propagate the identities of all target candidates from frame-to-frame. To tackle the problem of lacking ground-truth correspondences between distractor objects in visual tracking, we propose a training strategy that combines partial annotations with self-supervision. We conduct comprehensive experimental validation and analysis of our approach on several challenging datasets. Our tracker sets a new state-of-the-art on six benchmarks, achieving an AUC score of 67.1% on LaSOT [21] and a +5.8% absolute gain on the OxUvA long-term dataset [41]. The code and trained models are available at https://github.com/visionml/pytracking

It is widely acknowledged that single image super-resolution (SISR) methods would not perform well if the assumed degradation model deviates from those in real images. Although several degradation models take additional factors into consideration, such as blur, they are still not effective enough to cover the diverse degradations of real images. To address this issue, this paper proposes to design a more complex but practical degradation model that consists of randomly shuffled blur, downsampling and noise degradations. Specifically, the blur is approximated by two convolutions with isotropic and anisotropic Gaussian kernels; the downsampling is randomly chosen from nearest, bilinear and bicubic interpolations; the noise is synthesized by adding Gaussian noise with different noise levels, adopting JPEG compression with different quality factors, and generating processed camera sensor noise via reverse-forward camera image signal processing (ISP) pipeline model and RAW image noise model. To verify the effectiveness of the new degradation model, we have trained a deep blind ES-RGAN super-resolver and then applied it to super-resolve both synthetic and real images with diverse degradations. The experimental results demonstrate that the new degradation model can help to significantly improve the practicability of deep super-resolvers, thus providing a powerful alternative solution for real SISR applications.

Domain adaptation for semantic segmentation aims to improve the model performance in the presence of a distribution shift between source and target domain. Leveraging the supervision from auxiliary tasks (such as depth estimation) has the potential to heal this shift because many visual tasks are closely related to each other. However, such a supervision is not always available. In this work, we leverage the guidance from self-supervised depth estimation, which is available on both domains, to bridge the domain gap. On the one hand, we propose to explicitly learn the task feature correlation to strengthen the target semantic predictions with the help of target depth estimation. On the other hand, we use the depth prediction discrepancy from source and target depth decoders to approximate the pixel-wise adaptation difficulty. The adaptation difficulty, inferred from depth, is then used to refine the target semantic segmentation pseudo-labels. The proposed method can be easily implemented into existing segmentation frameworks. We demonstrate the effectiveness of our approach on the benchmark tasks SYNTHIA-to-Cityscapes and GTA-to-Cityscapes, on which we achieve the new state-of-the-art performance of 55.0% and 56.6%, respectively. Our code is available at https://qin.ee/corda.

We propose a deep reparametrization of the maximum a posteriori formulation commonly employed in multi-frame image restoration tasks. Our approach is derived by introducing a learned error metric and a latent representation of the target image, which transforms the MAP objective to a deep feature space. The deep reparametrization allows us to directly model the image formation process in the latent space, and to integrate learned image priors into the prediction. Our approach thereby leverages the advantages of deep learning, while also benefiting from the principled multi-frame fusion provided by the classical MAP formulation. We validate our approach through comprehensive experiments on burst denoising and burst super-resolution datasets. Our approach sets a new state-of-the-art for both tasks, demonstrating the generality and effectiveness of the proposed formulation.

The timeline of computer vision research is marked with advances in learning and utilizing efficient contextual representations. Most of them, however, are targeted at improving model performance on a single downstream task. We consider a multi-task environment for dense prediction tasks, represented by a common backbone and independent task-specific heads. Our goal is to find the most efficient way to refine each task prediction by capturing cross-task contexts dependent on tasks’ relations. We explore various attention-based contexts, such as global and local, in the multi-task setting and analyze their behavior when applied to refine each task independently. Empirical findings confirm that different source-target task pairs benefit from different context types. To automate the selection process, we propose an Adaptive Task-Relational Context (ATRC) module, which samples the pool of all available contexts for each task pair using neural architecture search and outputs the optimal configuration for deployment. Our method achieves state-of-the-art performance on two important multi-task benchmarks, namely NYUD-v2 and PASCAL-Context. The proposed ATRC has a low computational toll and can be used as a drop-in refinement module for any supervised multi-task architecture.

Autonomous navigation requires structured representation of the road network and instance-wise identification of the other traffic agents. Since the traffic scene is defined on the ground plane, this corresponds to scene understanding in the bird’s-eye-view (BEV). However, the onboard cameras of autonomous cars are customarily mounted horizontally for a better view of the surrounding, making this task very challenging. In this work, we study the problem of extracting a directed graph representing the local road network in BEV coordinates, from a single onboard camera image. Moreover, we show that the method can be extended to detect dynamic objects on the BEV plane. The semantics, locations, and orientations of the detected objects together with the road graph facilitates a comprehensive understanding of the scene. Such understanding becomes fundamental for the downstream tasks, such as path planning and navigation. We validate our approach against powerful baselines and show that our network achieves superior performance. We also demonstrate the effects of various design choices through ablation studies. Code: https://github.com/ybarancan/STSU

In this paper, we introduce a novel self-supervised visual representation learning method which understands both images and videos in a joint learning fashion. The proposed neural network architecture and objectives are designed to obtain two different Convolutional Neural Networks for solving visual recognition tasks in the domain of videos and images. Our method called Video/Image for Visual Contrastive Learning of Representation(Vi2CLR) uses unlabeled videos to exploit dynamic and static visual cues for self-supervised and instances similarity/dissimilarity learning. Vi2CLR optimization pipeline consists of visual clustering part and representation learning based on groups of similar positive instances within a cluster and negative ones from other clusters and learning visual clusters and their distances. We show how a joint self-supervised visual clustering and instance similarity learning with 2D (image) and 3D (video) CovNet encoders yields such robust and near to supervised learning performance.We extensively evaluate the method on downstream tasks like large scale action recognition, image and object classification on datasets like Kinetics, ImageNet, Pascal VOC’07 and UCF101 and achieve outstanding results compared to state-of-the-art self-supervised methods.

Many super-resolution (SR) models are optimized for high performance only and therefore lack efficiency due to large model complexity. As large models are often not practical in real-world applications, we investigate and propose novel loss functions, to enable SR with high perceptual quality from much more efficient models. The representative power for a given low-complexity generator network can only be fully leveraged by strong guidance towards the optimal set of parameters. We show that it is possible to improve the performance of a recently introduced efficient generator architecture solely with the application of our proposed loss functions. In particular, we use a Fourier space supervision loss for improved restoration of missing high-frequency (HF) content from the ground truth image and design a discriminator architecture working directly in the Fourier domain to better match the target HF distribution. We show that our losses’ direct emphasis on the frequencies in Fourier-space significantly boosts the perceptual image quality, while at the same time retaining high restoration quality in comparison to previously proposed loss functions for this task. The performance is further improved by utilizing a combination of spatial and frequency domain losses, as both representations provide complementary information during training. On top of that, the trained generator achieves comparable results with and is 2.4× and 48× faster than state-of-the-art perceptual SR methods RankSRGAN and SRFlow respectively.

Being able to learn dense semantic representations of images without supervision is an important problem in computer vision. However, despite its significance, this problem remains rather unexplored, with a few exceptions that considered unsupervised semantic segmentation on small-scale datasets with a narrow visual domain. In this paper, we make a first attempt to tackle the problem on datasets that have been traditionally utilized for the supervised case. To achieve this, we introduce a two-step framework that adopts a predetermined mid-level prior in a contrastive optimization objective to learn pixel embeddings. This marks a large deviation from existing works that relied on proxy tasks or end-to-end clustering. Additionally, we argue about the importance of having a prior that contains information about objects, or their parts, and discuss several possibilities to obtain such a prior in an unsupervised manner.Experimental evaluation shows that our method comes with key advantages over existing works. First, the learned pixel embeddings can be directly clustered in semantic groups using K-Means on PASCAL. Under the fully unsupervised setting, there is no precedent in solving the semantic segmentation task on such a challenging benchmark. Second, our representations can improve over strong baselines when transferred to new datasets, e.g. COCO and DAVIS. The code is available1.

One challenge of object recognition is to generalize to new domains, to more classes and/or to new modalities. This necessitates methods to combine and reuse existing datasets that may belong to different domains, have partial annotations, and/or have different data modalities. This paper formulates this as a multi-source domain adaptation and label unification problem, and proposes a novel method for it. Our method consists of a partially-supervised adaptation stage and a fully-supervised adaptation stage. In the former, partial knowledge is transferred from multiple source domains to the target domain and fused therein. Negative transfer between unmatching label spaces is mitigated via three new modules: domain attention, uncertainty maximization and attention-guided adversarial alignment. In the latter, knowledge is transferred in the unified label space after a label completion process with pseudolabels. Extensive experiments on three different tasks - image classification, 2D semantic image segmentation, and joint 2D-3D semantic segmentation - show that our method outperforms all competing methods significantly.

This work addresses the challenging task of LiDAR-based 3D object detection in foggy weather. Collecting and annotating data in such a scenario is very time, labor and cost intensive. In this paper, we tackle this problem by simulating physically accurate fog into clear-weather scenes, so that the abundant existing real datasets captured in clear weather can be repurposed for our task. Our contributions are twofold: 1) We develop a physically valid fog simulation method that is applicable to any LiDAR dataset. This unleashes the acquisition of large-scale foggy training data at no extra cost. These partially synthetic data can be used to improve the robustness of several perception methods, such as 3D object detection and tracking or simultaneous localization and mapping, on real foggy data. 2) Through extensive experiments with several state-of-the-art detection approaches, we show that our fog simulation can be leveraged to significantly improve the performance for 3D object detection in the presence of fog. Thus, we are the first to provide strong 3D object detection baselines on the Seeing Through Fog dataset. Our code is available at www.trace.ethz.ch/lidar fog simulation.

The state-of-the-art object detection and image classification methods can perform impressively on more than 9k classes. In contrast, the number of classes in semantic segmentation datasets is relatively limited. This is not surprising when the restrictions caused by the lack of labeled data and high computation demand for segmentation are considered. In this paper, we propose a novel training methodology to train and scale the existing semantic segmentation models for a large number of semantic classes without increasing the memory overhead. In our embedding-based scalable segmentation approach, we reduce the space complexity of the segmentation model’s output from O(C) to O(1), propose an approximation method for ground-truth class probability, and use it to compute cross-entropy loss. The proposed approach is general and can be adopted by any state-of-the-art segmentation model to gracefully scale it for any number of semantic classes with only one GPU. Our approach achieves similar, and in some cases, even better mIoU for Cityscapes, Pascal VOC, ADE20k, COCO-Stuff10k datasets when adopted to DeeplabV3+ model with different backbones. We demonstrate a clear benefit of our approach on a dataset with 1284 classes, bootstrapped from LVIS and COCO annotations, with almost three times better mIoU than the DeeplabV3+. Our source code is available at: https://github.com/shipra25jain/ESSNet.

We aim at improving the computational efficiency of graph convolutional networks (GCNs) for learning on point clouds. The basic graph convolution that is composed of a K-nearest neighbor (KNN) search and a multilayer perceptron (MLP) is examined. By mathematically analyzing the operations there, two findings to improve the efficiency of GCNs are obtained. (1) The local geometric structure information of 3D representations propagates smoothly across the GCN that relies on KNN search to gather neighborhood features. This motivates the simplification of multiple KNN searches in GCNs. (2) Shuffling the order of graph feature gathering and an MLP leads to equivalent or similar composite operations. Based on those findings, we optimize the computational procedure in GCNs. A series of experiments show that the optimized networks have reduced computational complexity, decreased memory consumption, and accelerated inference speed while maintaining comparable accuracy for learning on point clouds.

Current 3D object detection paradigms highly rely on extensive annotation efforts, which makes them not practical in many real-world industrial applications. Inspired by that a human driver can keep accumulating experiences from self-exploring the roads without any tutor’s guidance, we first step forwards to explore a simple yet effective self-supervised learning framework tailored for LiDAR-based 3D object detection. Although the self-supervised pipeline has achieved great success in 2D domain, the characteristic challenges (e.g., complex geometry structure and various 3D object views) encountered in the 3D domain hinder the direct adoption of existing techniques that often contrast the 2D augmented data or cluster single-view features. Here we present a novel self-supervised 3D Object detection framework that seamlessly integrates the geometry-aware contrast and clustering harmonization to lift the unsupervised 3D representation learning, named GCC-3D. First, GCC-3D introduces a Geometric-Aware Contrastive objective to learn spatial-sensitive local structure representation. This objective enforces the spatially close voxels to have high feature similarity. Second, a Pseudo-Instance Clustering harmonization mechanism is proposed to encourage that different views of pseudo-instances should have consistent similarities to clustering prototype centers. This module endows our model semantic discriminative capacity. Extensive experiments demonstrate our GCC-3D achieves significant performance improvement on data-efficient 3D object detection benchmarks (nuScenes and Waymo). Moreover, our GCC-3D framework can achieve state-of-the art performances on all popular 3D object detection benchmarks.

Normalizing flows have recently demonstrated promising results for low-level vision tasks. For image super-resolution (SR), it learns to predict diverse photo-realistic high-resolution (HR) images from the low-resolution (LR) image rather than learning a deterministic mapping. For image rescaling, it achieves high accuracy by jointly modelling the downscaling and upscaling processes. While existing approaches employ specialized techniques for these two tasks, we set out to unify them in a single formulation. In this paper, we propose the hierarchical conditional flow (HCFlow) as a unified framework for image SR and image rescaling. More specifically, HCFlow learns a bijective mapping between HR and LR image pairs by modelling the distribution of the LR image and the rest high-frequency component simultaneously. In particular, the high-frequency component is conditional on the LR image in a hierarchical manner. To further enhance the performance, other losses such as perceptual loss and GAN loss are combined with the commonly used negative log-likelihood loss in training. Extensive experiments on general image SR, face image SR and image rescaling have demonstrated that the proposed HCFlow achieves state-of-the-art performance in terms of both quantitative metrics and visual quality.

Existing blind image super-resolution (SR) methods mostly assume blur kernels are spatially invariant across the whole image. However, such an assumption is rarely applicable for real images whose blur kernels are usually spatially variant due to factors such as object motion and out-of-focus. Hence, existing blind SR methods would inevitably give rise to poor performance in real applications. To address this issue, this paper proposes a mutual affine network (MANet) for spatially variant kernel estimation. Specifically, MANet has two distinctive features. First, it has a moderate receptive field so as to keep the locality of degradation. Second, it involves a new mutual affine convolution (MAConv) layer that enhances feature expressiveness without increasing receptive field, model size and computation burden. This is made possible through exploiting channel interdependence, which applies each channel split with an affine transformation module whose input are the rest channel splits. Extensive experiments on synthetic and real images show that the proposed MANet not only performs favorably for both spatially variant and invariant kernel estimation, but also leads to state-of-the-art blind SR performance when combined with non-blind SR methods.

Level 5 autonomy for self-driving cars requires a robust visual perception system that can parse input images under any visual condition. However, existing semantic segmentation datasets are either dominated by images captured under normal conditions or are small in scale. To address this, we introduce ACDC, the Adverse Conditions Dataset with Correspondences for training and testing semantic segmentation methods on adverse visual conditions. ACDC consists of a large set of 4006 images which are equally distributed between four common adverse conditions: fog, nighttime, rain, and snow. Each adverse-condition image comes with a high-quality fine pixel-level semantic annotation, a corresponding image of the same scene taken under normal conditions, and a binary mask that distinguishes between intra-image regions of clear and uncertain semantic content. Thus, ACDC supports both standard semantic segmentation and the newly introduced uncertainty-aware semantic segmentation. A detailed empirical study demonstrates the challenges that the adverse domains of ACDC pose to state-of-the-art supervised and unsupervised approaches and indicates the value of our dataset in steering future progress in the field. Our dataset and benchmark are publicly available.

We introduce Task Switching Networks (TSNs), a task-conditioned architecture with a single unified encoder/decoder for efficient multi-task learning. Multiple tasks are performed by switching between them, performing one task at a time. TSNs have a constant number of parameters irrespective of the number of tasks. This scalable yet conceptually simple approach circumvents the overhead and intricacy of task-specific network components in existing works. In fact, we demonstrate for the first time that multi-tasking can be performed with a single task-conditioned decoder. We achieve this by learning task-specific conditioning parameters through a jointly trained task embedding network, encouraging constructive interaction between tasks. Experiments validate the effectiveness of our approach, achieving state-of-the-art results on two challenging multi-task benchmarks, PASCAL-Context and NYUD. Our analysis of the learned task embeddings further indicates a connection to task relationships studied in the recent literature.

The key challenge in learning dense correspondences lies in the lack of ground-truth matches for real image pairs. While photometric consistency losses provide unsupervised alternatives, they struggle with large appearance changes, which are ubiquitous in geometric and semantic matching tasks. Moreover, methods relying on synthetic training pairs often suffer from poor generalisation to real data.We propose Warp Consistency, an unsupervised learning objective for dense correspondence regression. Our objective is effective even in settings with large appearance and view-point changes. Given a pair of real images, we first construct an image triplet by applying a randomly sampled warp to one of the original images. We derive and analyze all flow-consistency constraints arising between the triplet. From our observations and empirical results, we design a general unsupervised objective employing two of the derived constraints. We validate our warp consistency loss by training three recent dense correspondence networks for the geometric and semantic matching tasks. Our approach sets a new stateof-the-art on several challenging benchmarks, including MegaDepth, RobotCar and TSS. Code and models are at github.com/PruneTruong/DenseMatching.

Current semantic segmentation methods focus only on mining "local" context, i.e., dependencies between pixels within individual images, by context-aggregation modules (e.g., dilated convolution, neural attention) or structure-aware optimization criteria (e.g., IoU-like loss). However, they ignore "global" context of the training data, i.e., rich semantic relations between pixels across different images. Inspired by recent advance in unsupervised contrastive representation learning, we propose a pixel-wise contrastive algorithm for semantic segmentation in the fully supervised setting. The core idea is to enforce pixel embeddings belonging to a same semantic class to be more similar than embeddings from different classes. It raises a pixel-wise metric learning paradigm for semantic segmentation, by explicitly exploring the structures of labeled pixels, which were rarely explored before. Our method can be effortlessly incorporated into existing segmentation frameworks without extra overhead during testing. We experimentally show that, with famous segmentation models (i.e., DeepLabV3, HRNet, OCR) and backbones (i.e., ResNet, HRNet), our method brings performance improvements across diverse datasets (i.e., Cityscapes, PASCAL-Context, COCO-Stuff, CamVid). We expect this work will encourage our community to rethink the current de facto training paradigm in semantic segmentation.

End-to-end approaches to autonomous driving commonly rely on expert demonstrations. Although humans are good drivers, they are not good coaches for end-to-end algorithms that demand dense on-policy supervision. On the contrary, automated experts that leverage privileged information can efficiently generate large scale on-policy and off-policy demonstrations. However, existing automated experts for urban driving make heavy use of hand-crafted rules and perform suboptimally even on driving simulators, where ground-truth information is available. To ad-dress these issues, we train a reinforcement learning expert that maps bird’s-eye view images to continuous low-level actions. While setting a new performance upper-bound on CARLA, our expert is also a better coach that provides in-formative supervision signals for imitation learning agents to learn from. Supervised by our reinforcement learning coach, a baseline end-to-end agent with monocular camera-input achieves expert-level performance. Our end-to-end agent achieves a 78% success rate while generalizing to a new town and new weather on the NoCrash-dense bench-mark and state-of-the-art performance on the more challenging CARLA LeaderBoard.

Segmenting objects in videos is a fundamental computer vision task. The current deep learning based paradigm offers a powerful, but data-hungry solution. However, current datasets are limited by the cost and human effort of annotating object masks in videos. This effectively limits the performance and generalization capabilities of existing video segmentation methods. To address this issue, we explore weaker form of bounding box annotations.We introduce a method for generating segmentation masks from per-frame bounding box annotations in videos. To this end, we propose a spatio-temporal aggregation module that effectively mines consistencies in the object and background appearance across multiple frames. We use our predicted accurate masks to train video object segmentation (VOS) networks for the tracking domain, where only manual bounding box annotations are available. The additional data provides substantially better generalization performance, leading to state-of-the-art results on standard tracking benchmarks. The code and models are available at https://github.com/visionml/pytracking.

Structure-from-Motion (SfM) using the frames of a video sequence can be a challenging task because there is a lot of redundant information, the computational time increases quadratically with the number of frames, there would be low-quality images (e.g., blurred frames) that can decrease the final quality of the reconstruction, etc. To overcome all these issues, we present a novel deep-learning architecture that is meant for speeding up SfM by selecting frames using predicted sub-sampling frequency. This architecture is general and can learn/distill the knowledge of any algorithm for selecting frames from a video for generating high-quality reconstructions. One key advantage is that we can run our architecture in real-time saving computations while keeping high-quality results.

We present a reinforcement learning-based solution to autonomously race on a miniature race car platform. We show that a policy that is trained purely in simulation using a relatively simple vehicle model, including model randomization, can be successfully transferred to the real robotic setup. We achieve this by using a novel policy output regularization approach and a lifted action space which enables smooth actions but still aggressive race car driving. We show that this regularized policy does outperform the Soft Actor Critic (SAC) baseline method, both in simulation and on the real car, but it is still outperformed by a Model Predictive Controller (MPC) state-of-the-art method. The refinement of the policy with three hours of real-world interaction data allows the reinforcement learning policy to achieve lap times similar to the MPC controller while reducing track constraint violations by 50%.

Autonomous agents interacting with the real world need to learn new concepts efficiently and reliably. This requires learning in a low-data regime, which is a highly challenging problem. We address this task by introducing a fast optimization-based meta-learning method for few-shot classification. It consists of an embedding network, providing a general representation of the image, and a base learner module. The latter learns a linear classifier during the inference through an unrolled optimization procedure. We design an inner learning objective composed of (i) a robust classification loss on the support set and (ii) an entropy loss, allowing transductive learning from unlabeled query samples. By employing an efficient initialization module and a Steepest Descent based optimization algorithm, our base learner predicts a powerful classifier within only a few iterations. Further, our strategy enables important aspects of the base learner objective to be learned during meta-training. To the best of our knowledge, this work is the first to integrate both induction and transduction into the base learner in an optimization-based meta-learning framework. We perform a comprehensive experimental analysis, demonstrating the speed and effectiveness of our approach on four few-shot classification datasets. The Code is available at https://github.com/4rdhendu/FIML.

In this paper, we propose a new neural architecture search (NAS) problem of Symmetric Positive Definite (SPD) manifold networks, aiming to automate the design of SPD neural architectures. To address this problem, we first introduce a geometrically rich and diverse SPD neural architecture search space for an efficient SPD cell design. Further, we model our new NAS problem with a one-shot training process of a single supernet. Based on the supernet modeling, we exploit a differentiable NAS algorithm on our relaxed continuous search space for SPD neural architecture search. Statistical evaluation of our method on drone, action, and emotion recognition tasks mostly provides better results than the state-of-the-art SPD networks and traditional NAS algorithms. Empirical results show that our algorithm excels in discovering better performing SPD network design and provides models that are more than three times lighter than searched by the state-of-the-art NAS algorithms.

We propose a novel neural network module that transforms an existing single-frame semantic segmentation model into a video semantic segmentation pipeline. In contrast to prior works, we strive towards a simple, fast, and general module that can be integrated into virtually any single-frame architecture. Our approach aggregates a rich representation of the semantic information in past frames into a memory module. Information stored in the memory is then accessed through an attention mechanism. In contrast to previous memory-based approaches, we propose a fast local attention layer, providing temporal appearance cues in the local region of prior frames. We further fuse these cues with an encoding of the current frame through a second attention-based module. The segmentation decoder processes the fused representation to predict the final semantic segmentation. We integrate our approach into two popular semantic segmentation networks: ERFNet and PSPNet. We observe an improvement in segmentation performance on Cityscapes by 1.7% and 2.1% in mIoU respectively, while increasing inference time of ERFNet by only 1.5ms. Source code is available at https://github.com/mattpfr/lmanet.

Forecasting the future behavior of all traffic agents in the vicinity is a key task to achieve safe and reliable autonomous driving systems. It is a challenging problem as agents adjust their behavior depending on their intentions, the others’ actions, and the road layout. In this paper, we propose Decoder Fusion RNN (DF-RNN), a recurrent, attention-based approach for motion forecasting. Our network is composed of a recurrent behavior encoder, an inter-agent multi-headed attention module, and a context-aware decoder. We design a map encoder that embeds polyline segments, combines them to create a graph structure, and merges their relevant parts with the agents’ embeddings. We fuse the encoded map information with further inter-agent interactions only inside the decoder and propose to use explicit training as a method to effectively utilize the information available. We demonstrate the efficacy of our method by testing it on the Argoverse motion forecasting dataset and show its state-of-the-art performance on the public benchmark.

Contrastive self-supervised learning has outperformed supervised pretraining on many downstream tasks like segmentation and object detection. However, current methods are still primarily applied to curated datasets like ImageNet. In this paper, we first study how biases in the dataset affect existing methods. Our results show that an approach like MoCo works surprisingly well across: (i) object- versus scene-centric, (ii) uniform versus long-tailed and (iii) general versus domain-specific datasets. Second, given the generality of the approach, we try to realize further gains with minor modifications. We show that learning additional invariances - through the use of multi-scale cropping, stronger augmentations and nearest neighbors - improves the representations. Finally, we observe that MoCo learns spatially structured representations when trained with a multi-crop strategy. The representations can be used for semantic segment retrieval and video instance segmentation without finetuning. Moreover, the results are on par with specialized models. We hope this work will serve as a useful study for other researchers.

While ground truth depth data remains hard to obtain, self-supervised monocular depth estimation methods enjoy growing attention. Much research in this area aims at improving loss functions or network architectures. Most works, however, do not leverage self-supervision to its full potential. They stick to the standard closed world train-test pipeline, assuming the network parameters to be fixed after the training is finished. Such an assumption does not allow to adapt to new scenes, whereas with self-supervision this becomes possible without extra annotations.In this paper, we propose a novel self-supervised Continuous Monocular Depth Adaptation method (CoMoDA), which adapts the pretrained model on a test video on the fly. As opposed to existing test-time refinement methods that use isolated frame triplets, we opt for continuous adaptation, making use of the previous experience from the same scene. We additionally augment the proposed procedure with the experience from the distant past, preventing the model from overfitting and thus forgetting already learnt information.We demonstrate that our method can be used for both intra- and cross-dataset adaptation. By adapting the model from train to test set of the Eigen split of KITTI, we achieve state-of-the-art depth estimation performance and surpass all existing methods using standard architectures. We also show that our method runs 15 times faster than existing test-time refinement methods. The code is available at https://github.com/Yevkuzn/CoMoDA.

In this paper, we propose an approach based on domain conditional normalization (DCN) for zero-pair image-to-image translation, i.e., translating between two domains which have no paired training data available but each have paired training data with a third domain. We employ a single generator which has an encoder-decoder structure and analyze different implementations of domain conditional normalization to obtain the desired target domain output. The validation benchmark uses RGB-depth pairs and RGB-semantic pairs for training and compares performance for the depth-semantic translation task. The proposed approaches improve in qualitative and quantitative terms over the compared methods, while using much fewer parameters.

Point cloud semantic segmentation plays an essential role in autonomous driving, providing vital information about drivable surfaces and nearby objects that can aid higher level tasks such as path planning and collision avoidance. While current 3D semantic segmentation networks focus on convolutional architectures that perform great for well represented classes, they show a significant drop in performance for underrepresented classes that share similar geometric features. We propose a novel Detection Aware 3D Semantic Segmentation (DASS) framework that explicitly leverages localization features from an auxiliary 3D object detection task. By utilizing multitask training, the shared feature representation of the network is guided to be aware of per class detection features that aid tackling the differentiation of geometrically similar classes. We additionally provide a pipeline that uses DASS to generate high recall proposals for existing 2-stage detectors and demonstrate that the added supervisory signal can be used to improve 3D orientation estimation capabilities. Extensive experiments on both the SemanticKITTI and KITTI object datasets show that DASS can improve 3D semantic segmentation results of geometrically similar classes up to 37.8% IoU in image FOV while maintaining high precision bird's-eye view (BEV) detection results.

Human emotions can be inferred from facial expressions. However, the annotations of facial expressions are often highly noisy in common emotion coding models, including categorical and dimensional ones. To reduce human labelling effort on multi-task labels, we introduce a new problem of facial emotion recognition with noisy multi-task annotations. For this new problem, we suggest a formulation from the point of joint distribution match view, which aims at learning more reliable correlations among raw facial images and multi-task labels, resulting in the reduction of noise influence. In our formulation, we exploit a new method to enable the emotion prediction and the joint distribution learning in a unified adversarial learning game. Evaluation throughout extensive experiments studies the real setups of the suggested new problem, as well as the clear superiority of the proposed method over the state-of-the-art competing methods on either the synthetic noisy labeled CIFAR-10 or practical noisy multi-task labeled RAF and AffectNet. The code is available at https://github.com/sanweiliti/noisyFER.

Abstract The role of robots in society keeps expanding, bringing with it the necessity of interacting and communicating with humans. In order to keep such interaction intuitive, we provide automatic wayfinding based on verbal navigational instructions. Our first contribution is the creation of a large-scale dataset with verbal navigation instructions. To this end, we have developed an interactive visual navigation environment based on Google Street View; we further design an annotation method to highlight mined anchor landmarks and local directions between them in order to help annotators formulate typical, human references to those. The annotation task was crowdsourced on the AMT platform, to construct a new Talk2Nav dataset with 10, 714 routes. Our second contribution is a new learning method. Inspired by spatial cognition research on the mental conceptualization of navigational instructions, we introduce a soft dual attention mechanism defined over the segmented language instructions to jointly extract two partial instructions—one for matching the next upcoming visual landmark and the other for matching the local directions to the next landmark. On the similar lines, we also introduce spatial memory scheme to encode the local directional transitions. Our work takes advantage of the advance in two lines of research: mental formalization of verbal navigational instructions and training neural network agents for automatic way finding. Extensive experiments show that our method significantly outperforms previous navigation methods. For demo video, dataset and code, please refer to our project page .

Abstract Reality TV shows have gained popularity, motivating many production houses to bring new variants for us to watch. Compared to traditional TV shows, reality TV shows have spontaneous unscripted footage. Computer vision techniques could partially replace the manual labour needed to record and process this spontaneity. However, automated real-world video recording and editing is a challenging topic. In this paper, we propose a system that utilises state-of-the-art video and audio processing algorithms to, on the one hand, automatically steer cameras, replacing camera operators and on the other hand, detect all audiovisual action cues in the recorded video, to ease the job of the film editor. This publication has hence two main contributions. The first, automating the steering of multiple Pan-Tilt-Zoom PTZ cameras to take aesthetically pleasing medium shots of all the people present. These shots need to comply with the cinematographic rules and are based on the poses acquired by a pose detector. Secondly, when a huge amount of audio-visual data has been collected, it becomes labour intensive for a human editor retrieve the relevant fragments. As a second contribution, we combine state-of-the-art audio and video processing techniques for sound activity detection, action recognition, face recognition, and pose detection to decrease the required manual labour during and after recording. These techniques used during post-processing produce meta-data allowing for footage filtering, decreasing the search space. We extended our system further by producing timelines uniting generated meta-data, allowing the editor to have a quick overview. We evaluated our system on three in-the-wild reality TV recording sessions of 24 hours (× 8 cameras) each taken in real households.

Visual tracking is fundamentally the problem of regressing the state of the target in each video frame. While significant progress has been achieved, trackers are still prone to failures and inaccuracies. It is therefore crucial to represent the uncertainty in the target estimation. Although current prominent paradigms rely on estimating a state-dependent confidence score, this value lacks a clear probabilistic interpretation, complicating its use. In this work, we therefore propose a probabilistic regression formulation and apply it to tracking. Our network predicts the conditional probability density of the target state given an input image. Crucially, our formulation is capable of modeling label noise stemming from inaccurate annotations and ambiguities in the task. The regression network is trained by minimizing the Kullback-Leibler divergence. When applied for tracking, our formulation not only allows a probabilistic representation of the output, but also substantially improves the performance. Our tracker sets a new state-of-the-art on six datasets, achieving 59.8% AUC on LaSOT and 75.8% Success on TrackingNet. The code and models are available at https://github.com/visionml/pytracking.

In this paper, we analyze two popular network compression techniques, i.e. filter pruning and low-rank decomposition, in a unified sense. By simply changing the way the sparsity regularization is enforced, filter pruning and low-rank decomposition can be derived accordingly. This provides another flexible choice for network compression because the techniques complement each other. For example, in popular network architectures with shortcut connections (e.g. ResNet), filter pruning cannot deal with the last convolutional layer in a ResBlock while the low-rank decomposition methods can. In addition, we propose to compress the whole network jointly instead of in a layer-wise manner. Our approach proves its potential as it compares favorably to the state-of-the-art on several benchmarks. Code is available at https://github.com/ofsoundof/group_sparsity.

We leverage the powerful lossy image compression algorithm BPG to build a lossless image compression system. Specifically, the original image is first decomposed into the lossy reconstruction obtained after compressing it with BPG and the corresponding residual. We then model the distribution of the residual with a convolutional neural network-based probabilistic model that is conditioned on the BPG reconstruction, and combine it with entropy coding to losslessly encode the residual. Finally, the image is stored using the concatenation of the bitstreams produced by BPG and the learned residual coder. The resulting compression system achieves state-of-the-art performance in learned lossless full-resolution image compression, outperforming previous learned approaches as well as PNG, WebP, and JPEG2000.

Humans perceive the 3D world as a set of distinct objects that are characterized by various low-level (geometry, reflectance) and high-level (connectivity, adjacency, symmetry) properties. Recent methods based on convolutional neural networks (CNNs) demonstrated impressive progress in 3D reconstruction, even when using a single 2D image as input. However, the majority of these methods focuses on recovering the local 3D geometry of an object without considering its part-based decomposition or relations between parts. We address this challenging problem by proposing a novel formulation that allows to jointly recover the geometry of a 3D object as a set of primitives as well as their latent hierarchical structure without part-level supervision. Our model recovers the higher level structural decomposition of various objects in the form of a binary tree of primitives, where simple parts are represented with fewer primitives and more complex parts are modeled with more components. Our experiments on the ShapeNet and D-FAUST datasets demonstrate that considering the organization of parts indeed facilitates reasoning about 3D geometry.

In this paper, we propose a Hierarchical Learned Video Compression (HLVC) method with three hierarchical quality layers and a recurrent enhancement network. The frames in the first layer are compressed by an image compression method with the highest quality. Using these frames as references, we propose the Bi-Directional Deep Compression (BDDC) network to compress the second layer with relatively high quality. Then, the third layer frames are compressed with the lowest quality, by the proposed Single Motion Deep Compression (SMDC) network, which adopts a single motion map to estimate the motions of multiple frames, thus saving bits for motion information. In our deep decoder, we develop the Weighted Recurrent Quality Enhancement (WRQE) network, which takes both compressed frames and the bit stream as inputs. In the recurrent cell of WRQE, the memory and update signal are weighted by quality features to reasonably leverage multi-frame information for enhancement. In our HLVC approach, the hierarchical quality benefits the coding efficiency, since the high quality information facilitates the compression and enhancement of low quality frames at encoder and decoder sides, respectively. Finally, the experiments validate that our HLVC approach advances the state-of-the-art of deep video compression methods, and outperforms the "Low-Delay P (LDP) very fast" mode of x265 in terms of both PSNR and MS-SSIM. The project page is at https://github.com/RenYang-home/HLVC.

Learning-based single image super-resolution (SISR) methods are continuously showing superior effectiveness and efficiency over traditional model-based methods, largely due to the end-to-end training. However, different from model-based methods that can handle the SISR problem with different scale factors, blur kernels and noise levels under a unified MAP (maximum a posteriori) framework, learning-based methods generally lack such flexibility. To address this issue, this paper proposes an end-to-end trainable unfolding network which leverages both learningbased methods and model-based methods. Specifically, by unfolding the MAP inference via a half-quadratic splitting algorithm, a fixed number of iterations consisting of alternately solving a data subproblem and a prior subproblem can be obtained. The two subproblems then can be solved with neural modules, resulting in an end-to-end trainable, iterative network. As a result, the proposed network inherits the flexibility of model-based methods to super-resolve blurry, noisy images for different scale factors via a single model, while maintaining the advantages of learning-based methods. Extensive experiments demonstrate the superiority of the proposed deep unfolding network in terms of flexibility, effectiveness and also generalizability.

We address the problem of unpaired geometric image-to-image translation. Rather than transferring the style of an image as a whole, our goal is to translate the geometry of an object while preserving its appearance. Our model is trained without the need for paired images. It performs all steps of the shape transfer within a single model and without additional post-processing stages. Experiments on clothing-based datasets show the effectiveness of the proposed method.

We introduce T-Basis, a novel concept for a compact representation of a set of tensors, each of an arbitrary shape, which is often seen in Neural Networks. Each of the tensors in the set is modeled using Tensor Rings, though the concept applies to other Tensor Networks. Owing its name to the T-shape of nodes in diagram notation of Tensor Rings, T-Basis is simply a list of equally shaped three-dimensional tensors, used to represent Tensor Ring nodes. Such representation allows us to parameterize the tensor set with a small number of parameters (coefficients of the T-Basis tensors), scaling logarithmically with each tensor’s size in the set and linearly with the dimensionality of T-Basis. We evaluate the proposed approach on the task of neural network compression and demonstrate that it reaches high compression rates at acceptable performance drops. Finally, we analyze memory and operation requirements of the compressed networks and conclude that T-Basis networks are equally well suited for training and inference in resource-constrained environments and usage on the edge devices. Project website: obukhov.ai/tbasis.

This paper investigates how end-to-end driving models can be improved to drive more accurately and human-like. To tackle the first issue we exploit semantic and visual maps from HERE Technologies and augment the existing Drive360 dataset with such. The maps are used in an attention mechanism that promotes segmentation confidence masks, thus focusing the network on semantic classes in the image that are important for the current driving situation. Human-like driving is achieved using adversarial learning, by not only minimizing the imitation loss with respect to the human driver but by further defining a discriminator, that forces the driving model to produce action sequences that are human-like. Our models are trained and evaluated on the Drive360 + HERE dataset, which features 60 hours and 3000 km of real-world driving data. Extensive experiments show that our driving models are more accurate and behave more human-like than previous methods.

In this paper, we explore the possibility of achieving a more accurate depth estimation by fusing monocular images and Radar points using a deep neural network. We give a comprehensive study of the fusion between RGB images and Radar measurements from different aspects and proposed a working solution based on the observations. We find that the noise existing in Radar measurements is one of the main key reasons that prevents one from applying the existing fusion methods developed for LiDAR data and images to the new fusion problem between Radar data and images. The experiments are conducted on the nuScenes dataset, which is one of the first datasets which features Camera, Radar, and LiDAR recordings in diverse scenes and weather conditions. Extensive experiments demonstrate that our method outperforms existing fusion methods. We also provide detailed ablation studies to show the effectiveness of each component in our method.

This work studies the problem of predicting the sequence of future actions for surrounding vehicles in real-world driving scenarios. To this aim, we make three main contributions. The first contribution is an automatic method to convert the trajectories recorded in real-world driving scenarios to action sequences with the help of HD maps. The method enables automatic dataset creation for this task from large-scale driving data. Our second contribution lies in applying the method to the well-known traffic agent tracking and prediction dataset Argoverse, resulting in 228,000 action sequences. Additionally, 2,245 action sequences were manually annotated for testing. The third contribution is to propose a novel action sequence prediction method by integrating past positions and velocities of the traffic agents, map information and social context into a single end-to-end trainable neural network. Our experiments prove the merit of the data creation method and the value of the created dataset - prediction performance improves consistently with the size of the dataset and shows that our action prediction method outperforms comparing models.

Localization by image retrieval is inexpensive and scalable due to simple mapping and matching techniques. Such localization, however, depends upon the quality of image features often obtained using Contrastive learning frameworks. Most contrastive learning strategies opt for features to distinguish different classes. In the context of localization, however, there is no natural definition of classes. Therefore, images are usually artificially separated into positive and negative classes, with respect to the chosen anchor images, based on some geometric proximity measure. In this paper, we show why such divisions are problematic for learning localization features. We argue that any artificial division based on some proximity measure is undesirable, due to the inherently ambiguous supervision for images near proximity threshold. To this end, we propose a novel technique that uses soft positive/negative assignments of images for contrastive learning, avoiding the aforementioned problem. Our soft assignment makes a gradual distinction between close and far images in both geometric and feature spaces. Experiments on four large-scale benchmark datasets demonstrate the superiority of the proposed soft contrastive learning over the state-of-the-art method for retrieval-based visual localization.

The feature correlation layer serves as a key neural network module in numerous computer vision problems that involve dense correspondences between image pairs. It predicts a correspondence volume by evaluating dense scalar products between feature vectors extracted from pairs of locations in two images. However, this point-to-point feature comparison is insufficient when disambiguating multiple similar regions in an image, severely affecting the performance of the end task. We propose GOCor, a fully differentiable dense matching module, acting as a direct replacement to the feature correlation layer. The correspondence volume generated by our module is the result of an internal optimization procedure that explicitly accounts for similar regions in the scene. Moreover, our approach is capable of effectively learning spatial matching priors to resolve further matching ambiguities. We analyze our GOCor module in extensive ablative experiments. When integrated into state-of-the-art networks, our approach significantly outperforms the feature correlation layer for the tasks of geometric matching, optical flow, and dense semantic matching. The code and trained models will be made available at github.com/PruneTruong/GOCor.

Monocular depth estimation has gained increasing attention in recent years, and various techniques have been proposed to tackle this problem. In this work, we aim to provide a comprehensive study on the techniques widely used in monocular depth estimation, and examine their individual influence on the performance. More specifically, we provide a study on: 1) network architectures, including different combinations of encoders/decoders. 2) supervision losses, including fully supervised losses and self-supervised losses and 3) other practices such as input resolution. The experiments are conducted on two commonly used public datasets, KITTI and NYU Depth v2. We also provide an analysis on the errors produced by different models, to reveal the limitations of current methods. Furthermore, by a careful redesign, we present a model for depth estimation, which achieves competitive performance on KITTI and state-of-the-art performance on NYU Depth v2. Our code is publicly available at https://github.com/zenithfang/supervised_dispnet.

This paper tackles the problem of real-time semantic segmentation of high definition videos using a hybrid GPU-CPU approach. We propose an Efficient Video Segmentation (EVS) pipeline that combines:(i)On the CPU, a very fast optical flow method, that is used to exploit the temporal aspect of the video and propagate semantic information from one frame to the next. It runs in parallel with the GPU.(ii)On the GPU, two Convolutional Neural Networks: A main segmentation network that is used to predict dense semantic labels from scratch, and a Refiner that is designed to improve predictions from previous frames with the help of a fast Inconsistencies Attention Module (IAM). The latter can identify regions that cannot be propagated accurately.We suggest several operating points depending on the desired frame rate and accuracy. Our pipeline achieves accuracy levels competitive to the existing real-time methods for semantic image segmentation (mIoU above 60%), while achieving much higher frame rates. On the popular Cityscapes dataset with high resolution frames (2048 × 1024), the proposed operating points range from 80 to 1000 Hz on a single GPU and CPU.

The depth images acquired by consumer depth sensors (e.g., Kinect and ToF) usually are of low resolution and insufficient quality. One natural solution is to incorporate a high resolution RGB camera and exploit the statistical correlation of its data and depth. In recent years, both optimization-based and learning-based approaches have been proposed to deal with the guided depth reconstruction problems. In this paper, we introduce a weighted analysis sparse representation (WASR) model for guided depth image enhancement, which can be considered a generalized formulation of a wide range of previous optimization-based models. We unfold the optimization by the WASR model and conduct guided depth reconstruction with dynamically changed stage-wise operations. Such a guidance strategy enables us to dynamically adjust the stage-wise operations that update the depth image, thus improving the reconstruction quality and speed. To learn the stage-wise operations in a task-driven manner, we propose two parameterizations and their corresponding methods: dynamic guidance with Gaussian RBF nonlinearity parameterization (DG-RBF) and dynamic guidance with CNN nonlinearity parameterization (DG-CNN). The network structures of the proposed DG-RBF and DG-CNN methods are designed with the the objective function of our WASR model in mind and the optimal network parameters are learned from paired training data. Such optimization-inspired network architectures enable our models to leverage the previous expertise as well as take benefit from training data. The effectiveness is validated for guided depth image super-resolution and for realistic depth image reconstruction tasks using standard benchmarks. Our DG-RBF and DG-CNN methods achieve the best quantitative results (RMSE) and better visual quality than the state-of-the-art approaches at the time of writing. The code is available at https://github.com/ShuhangGu/GuidedDepthSR.

In real life, group activity recognition plays a significant and fundamental role in a variety of applications, e.g. sports video analysis, abnormal behavior detection, and intelligent surveillance. In a complex dynamic scene, a crucial yet challenging issue is how to better model the spatio-temporal contextual information and inter-person relationship. In this paper, we present a novel attentive semantic recurrent neural network (RNN), namely, stagNet, for understanding group activities and individual actions in videos, by combining the spatio-temporal attention mechanism and semantic graph modeling. Specifically, a structured semantic graph is explicitly modeled to express the spatial contextual content of the whole scene, which is further incorporated with the temporal factor through structural-RNN. By virtue of the “factor sharing” and “message passing” mechanisms, our stagNet is capable of extracting discriminative and informative spatio-temporal representations and capturing inter-person relationships. Moreover, we adopt a spatio-temporal attention model to focus on key persons/frames for improved recognition performance. Besides, a body-region attention and a global-part feature pooling strategy are devised for individual action recognition. In experiments, four widely-used public datasets are adopted for performance evaluation, and the extensive results demonstrate the superiority and effectiveness of our method.

Generative modeling over natural images is one of the most fundamental machine learning problems. However, few modern generative models, including Wasserstein Generative Adversarial Nets (WGANs), are studied on manifold-valued images that are frequently encountered in real-world applications. To fill the gap, this paper first formulates the problem of generating manifold-valued images and exploits three typical instances: hue-saturation-value (HSV) color image generation, chromaticity-brightness (CB) color image generation, and diffusion-tensor (DT) image generation. For the proposed generative modeling problem, we then introduce a theorem of optimal transport to derive a new Wasserstein distance of data distributions on complete manifolds, enabling us to achieve a tractable objective under the WGAN framework. In addition, we recommend three benchmark datasets that are CIFAR-10 HSV/CB color images, ImageNet HSV/CB color images, UCL DT image datasets. On the three datasets, we experimentally demonstrate the proposed manifold-aware WGAN model can generate more plausible manifold-valued images than its competitors.

As an alternative to manual pixel-wise annotation, synthetic data has been increasingly used for training semantic segmentation models. Such synthetic images and semantic labels can be easily generated from virtual 3D environments. In this work, we propose an approach to cross-domain semantic segmentation with the auxiliary geometric information, which can also be easily obtained from virtual environments. The geometric information is utilized on two levels for reducing domain shift: on the input level, we augment the standard image translation network with the geometric information to translate synthetic images into realistic style; on the output level, we build a task network which simultaneously performs semantic segmentation and depth estimation. Meanwhile, adversarial training is applied on the joint output space to preserve the correlation between semantics and depth. The proposed approach is validated on two pairs of synthetic to real dataset: from Virtual KITTI to KITTI, and from SYNTHIA to Cityscapes, where we achieve a clear performance gain compared to the baselines and various competing methods, demonstrating the effectiveness of the geometric information for cross-domain semantic segmentation.

In this work, we present a domain flow generation(DLOW) model to bridge two different domains by generating a continuous sequence of intermediate domains flowing from one domain to the other. The benefits of our DLOW model are two-fold. First, it is able to transfer source images into different styles in the intermediate domains. The transferred images smoothly bridge the gap between source and target domains, thus easing the domain adaptation task. Second, when multiple target domains are provided for training, our DLOW model is also able to generate new styles of images that are unseen in the training data. We implement our DLOW model based on CycleGAN. A domainness variable is introduced to guide the model to generate the desired intermediate domain images. In the inference phase, a flow of various styles of images can be obtained by varying the domainness variable. We demonstrate the effectiveness of our model for both cross-domain semantic segmentation and the style generalization tasks on benchmark datasets. Our implementation is available at https://github.com/ETHRuiGong/DLOW .

We tackle the problem of retrieving high-resolution (HR) texture maps of objects that are captured from multiple view points. In the multi-view case, model-based super-resolution (SR) methods have been recently proved to recover high quality texture maps. On the other hand, the advent of deep learning-based methods has already a significant impact on the problem of video and image SR. Yet, a deep learning-based approach to super-resolve the appearance of 3D objects is still missing. The main limitation of exploiting the power of deep learning techniques in the multi-view case is the lack of data. We introduce a 3D appearance SR (3DASR) dataset based on the existing ETH3D [42], SyB3R [31], MiddleBury, and our Collection of 3D scenes from TUM [21], Fountain [51] and Relief [53]. We provide the high- and low-resolution texture maps, the 3D geometric model, images and projection matrices. We exploit the power of 2D learning-based SR methods and design networks suitable for the 3D multi-view case. We incorporate the geometric information by introducing normal maps and further improve the learning process. Experimental results demonstrate that our proposed networks successfully incorporate the 3D geometric information and super-resolve the texture maps.

We propose the first practical learned lossless image compression system, L3C, and show that it outperforms the popular engineered codecs, PNG, WebP and JPEG 2000. At the core of our method is a fully parallelizable hierarchical probabilistic model for adaptive entropy coding which is optimized end-to-end for the compression task. In contrast to recent autoregressive discrete probabilistic models such as PixelCNN, our method i) models the image distribution jointly with learned auxiliary representations instead of exclusively modeling the image distribution in RGB space, and ii) only requires three forward-passes to predict all pixel probabilities instead of one for each pixel. As a result, L3C obtains over two orders of magnitude speedups when sampling compared to the fastest PixelCNN variant (Multiscale-PixelCNN). Furthermore, we find that learning the auxiliary representation is crucial and outperforms predefined auxiliary representations such as an RGB pyramid significantly.

Current state-of-the-art instance segmentation methods are not suited for real-time applications like autonomous driving, which require fast execution times at high accuracy. Although the currently dominant proposal-based methods have high accuracy, they are slow and generate masks at a fixed and low resolution. Proposal-free methods, by contrast, can generate masks at high resolution and are often faster, but fail to reach the same accuracy as the proposal-based methods. In this work we propose a new clustering loss function for proposal-free instance segmentation. The loss function pulls the spatial embeddings of pixels belonging to the same instance together and jointly learns an instance-specific clustering bandwidth, maximizing the intersection-over-union of the resulting instance mask. When combined with a fast architecture, the network can perform instance segmentation in real-time while maintaining a high accuracy. We evaluate our method on the challenging Cityscapes benchmark and achieve top results (5% improvement over Mask R-CNN) at more than 10 fps on 2MP images.

Consensus maximization is a key strategy in 3D vision for robust geometric model estimation from measurements with outliers. Generic methods for consensus maximization, such as Random Sampling and Consensus (RANSAC), have played a tremendous role in the success of 3D vision, in spite of the ubiquity of outliers. However, replicating the same generic behaviour in a deeply learned architecture, using supervised approaches, has proven to be difficult. In that context, unsupervised methods have a huge potential to adapt to any unseen data distribution, and therefore are highly desirable. In this paper, we propose for the first time an unsupervised learning framework for consensus maximization, in the context of solving 3D vision problems. For that purpose, we establish a relationship between inlier measurements, represented by an ideal of inlier set, and the subspace of polynomials representing the space of target transformations. Using this relationship, we derive a constraint that must be satisfied by the sought inlier set. This constraint can be tested without knowing the transformation parameters, therefore allows us to efficiently define the geometric model fitting cost. This model fitting cost is used as a supervisory signal for learning consensus maximization, where the learning process seeks for the largest measurement set that minimizes the proposed model fitting cost. Using our method, we solve a diverse set of 3D vision problems, including 3D-3D matching, non-rigid 3D shape matching with piece-wise rigidity and image-to-image matching. Despite being unsupervised, our method outperforms RANSAC in all three tasks for several datasets.

Building on progress in feature representations for image retrieval, image-based localization has seen a surge of research interest. Image-based localization has the advantage of being inexpensive and efficient, often avoiding the use of 3D metric maps altogether. That said, the need to maintain a large amount of reference images as an effective support of localization in a scene, nonetheless calls for them to be organized in a map structure of some kind. The problem of localization often arises as part of a navigation process. We are, therefore, interested in summarizing the reference images as a set of landmarks, which meet the requirements for image-based navigation. A contribution of this paper is to formulate such a set of requirements for the two sub-tasks involved: compact map construction and accurate self localization. These requirements are then exploited for compact map representation and accurate self-localization, using the framework of a network flow problem. During this process, we formulate the map construction and self-localization problems as convex quadratic and second-order cone programs, respectively. We evaluate our methods on publicly available indoor and outdoor datasets, where they outperform existing methods significantly.

We present a self-supervision method for 3D hand pose estimation from depth maps. We begin with a neural network initialized with synthesized data and fine-tune it on real but unlabelled depth maps by minimizing a set of data-fitting terms. By approximating the hand surface with a set of spheres, we design a differentiable hand renderer to align estimates by comparing the rendered and input depth maps. In addition, we place a set of priors including a data-driven term to further regulate the estimate's kinematic feasibility. Our method makes highly accurate estimates comparable to current supervised methods which require large amounts of labelled training samples, thereby advancing state-of-the-art in unsupervised learning for hand pose estimation.

In generative modeling, the Wasserstein distance (WD) has emerged as a useful metric to measure the discrepancy between generated and real data distributions. Unfortunately, it is challenging to approximate the WD of high-dimensional distributions. In contrast, the sliced Wasserstein distance (SWD) factorizes high-dimensional distributions into their multiple one-dimensional marginal distributions and is thus easier to approximate. In this paper, we introduce novel approximations of the primal and dual SWD. Instead of using a large number of random projections, as it is done by conventional SWD approximation methods, we propose to approximate SWDs with a small number of parameterized orthogonal projections in an end-to-end deep learning fashion. As concrete applications of our SWD approximations, we design two types of differentiable SWD blocks to equip modern generative frameworks---Auto-Encoders (AE) and Generative Adversarial Networks (GAN). In the experiments, we not only show the superiority of the proposed generative models on standard image synthesis benchmarks, but also demonstrate the state-of-the-art performance on challenging high resolution image and video generation in an unsupervised manner.

Thierry Deruyttere, Simon Vandenhende, Dusan Grujicic, Luc Van Gool, Marie-Francine Moens. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

We present a learned image compression system based on GANs, operating at extremely low bitrates. Our proposed framework combines an encoder, decoder/generator and a multi-scale discriminator, which we train jointly for a generative learned compression objective. The model synthesizes details it cannot afford to store, obtaining visually pleasing results at bitrates where previous methods fail and show strong artifacts. Furthermore, if a semantic label map of the original image is available, our method can fully synthesize unimportant regions in the decoded image such as streets and trees from the label map, proportionally reducing the storage cost. A user study confirms that for low bitrates, our approach is preferred to state-of-the-art methods, even when they use more than double the bits.

The current strive towards end-to-end trainable computer vision systems imposes major challenges for the task of visual tracking. In contrast to most other vision problems, tracking requires the learning of a robust target-specific appearance model online, during the inference stage. To be end-to-end trainable, the online learning of the target model thus needs to be embedded in the tracking architecture itself. Due to the imposed challenges, the popular Siamese paradigm simply predicts a target feature template, while ignoring the background appearance information during inference. Consequently, the predicted model possesses limited target-background discriminability. We develop an end-to-end tracking architecture, capable of fully exploiting both target and background appearance information for target model prediction. Our architecture is derived from a discriminative learning loss by designing a dedicated optimization process that is capable of predicting a powerful model in only a few iterations. Furthermore, our approach is able to learn key aspects of the discriminative loss itself. The proposed tracker sets a new state-of-the-art on 6 tracking benchmarks, achieving an EAO score of 0.440 on VOT2018, while running at over 40 FPS. The code and models are available at https://github.com/visionml/pytracking.

In this paper, we are interested in self-supervised learning the motion cues in videos using dynamic motion filters for a better motion representation to finally boost human action recognition in particular. Thus far, the vision community has focused on spatio-temporal approaches using standard filters, rather we here propose dynamic filters that adaptively learn the video-specific internal motion representation by predicting the short-term future frames. We name this new motion representation, as dynamic motion representation (DMR) and is embedded inside of 3D convolutional network as a new layer, which captures the visual appearance and motion dynamics throughout entire video clip via end-to-end network learning. Simultaneously, we utilize these motion representation to enrich video classification. We have designed the frame prediction task as an auxiliary task to empower the classification problem. With these overall objectives, to this end, we introduce a novel unified spatio-temporal 3D-CNN architecture (DynamoNet) that jointly optimizes the video classification and learning motion representation by predicting future frames as a multi-task learning problem. We conduct experiments on challenging human action datasets: Kinetics 400, UCF101, HMDB51. The experiments using the proposed DynamoNet show promising results on all the datasets.

During the past years, tremendous advances in image restoration tasks have been achieved using highly complex neural networks. Despite their good restoration performance, the heavy computational burden hinders the deployment of these networks on constrained devices, \eg smart phones and consumer electronic products. To tackle this problem, we propose a self-guided network (SGN), which adopts a top-down self-guidance architecture to better exploit image multi-scale information. SGN directly generates multi-resolution inputs with the shuffling operation. Large-scale contextual information extracted at low resolution is gradually propagated into the higher resolution sub-networks to guide the feature extraction processes at these scales. Such a self-guidance strategy enables SGN to efficiently incorporate multi-scale information and extract good local features to recover noisy images. We validate the effectiveness of SGN through extensive experiments. The experimental results demonstrate that SGN greatly improves the memory and runtime efficiency over state-of-the-art efficient methods, without trading off PSNR accuracy.

Tremendous advances in image restoration tasks such as denoising and super-resolution have been achieved using neural networks. Such approaches generally employ very deep architectures, large number of parameters, large receptive fields and high nonlinear modeling capacity. In order to obtain efficient and fast image restoration networks one should improve upon the above mentioned requirements. In this paper we propose a novel activation function, the multi-bin trainable linear unit (MTLU), for increasing the nonlinear modeling capacity together with lighter and shallower networks. We validate the proposed fast image restoration networks for image denoising (FDnet) and super-resolution (FSRnet) on standard benchmarks. We achieve large improvements in both memory and runtime over current state-of-the-art for comparable or better PSNR accuracies.

Convolutional neural networks (CNNs) based solutions have achieved state-of-the-art performances for many computer vision tasks, including classification and super-resolution of images. Usually the success of these methods comes with a cost of millions of parameters due to stacking deep convolutional layers. Moreover, quite a large number of filters are also used for a single convolutional layer, which exaggerates the parameter burden of current methods. Thus, in this paper, we try to reduce the number of parameters of CNNs by learning a basis of the filters in convolutional layers. For the forward pass, the learned basis is used to approximate the original filters and then used as parameters for the convolutional layers. We validate our proposed solution for multiple CNN architectures on image classification and image super-resolution benchmarks and compare favorably to the existing state-of-the-art in terms of reduction of parameters and preservation of accuracy. Code is available at https://github.com/ofsoundof/learning_filter_basis.

In this paper, we formulate a generic non-minimal solver using the existing tools of Polynomials Optimization Problems (POP) from computational algebraic geometry. The proposed method exploits the well known Shor's or Lasserre's relaxations, whose theoretical aspects are also discussed. Notably, we further exploit the POP formulation of non-minimal solver also for the generic consensus maximization problems in 3D vision. Our framework is simple and straightforward to implement, which is also supported by three diverse applications in 3D vision, namely rigid body transformation estimation, Non-Rigid Structure-from-Motion (NRSfM), and camera autocalibration. In all three cases, both non-minimal and consensus maximization are tested, which are also compared against the state-of-the-art methods. Our results are competitive to the compared methods, and are also coherent with our theoretical analysis. The main contribution of this paper is the claim that a good approximate solution for many polynomial problems involved in 3D vision can be obtained using the existing theory of numerical computational algebra. This claim leads us to reason about why many relaxed methods in 3D vision behave so well? And also allows us to offer a generic relaxed solver in a rather straightforward way. We further show that the convex relaxation of these polynomials can easily be used for maximizing consensus in a deterministic manner. We support our claim using several experiments for aforementioned three diverse problems in 3D vision.

Most progress in semantic segmentation reports on daytime images taken under favorable illumination conditions. We instead address the problem of semantic segmentation of nighttime images and improve the state-of-the-art, by adapting daytime models to nighttime without using nighttime annotations. Moreover, we design a new evaluation framework to address the substantial uncertainty of semantics in nighttime images. Our central contributions are: 1) a curriculum framework to gradually adapt semantic segmentation models from day to night via labeled synthetic images and unlabeled real images, both for progressively darker times of day, which exploits cross-time-of-day correspondences for the real images to guide the inference of their labels; 2) a novel uncertainty-aware annotation and evaluation framework and metric for semantic segmentation, designed for adverse conditions and including image regions beyond human recognition capability in the evaluation in a principled fashion; 3) the Dark Zurich dataset, which comprises 2416 unlabeled nighttime and 2920 unlabeled twilight images with correspondences to their daytime counterparts plus a set of 151 nighttime images with fine pixel-level annotations created with our protocol, which serves as a first benchmark to perform our novel evaluation. Experiments show that our guided curriculum adaptation significantly outperforms state-of-the-art methods on real nighttime sets both for standard metrics and our uncertainty-aware metric. Furthermore, our uncertainty-aware evaluation reveals that selective invalidation of predictions can lead to better results on data with ambiguous content such as our nighttime benchmark and profit safety-oriented applications which involve invalid inputs.

In this work, we propose a simple yet effective semi-supervised learning approach called Augmented Distribution Alignment. We reveal that an essential sampling bias exists in semi-supervised learning due to the limited number of labeled samples, which often leads to a considerable empirical distribution mismatch between labeled data and unlabeled data. To this end, we propose to align the empirical distributions of labeled and unlabeled data to alleviate the bias. On one hand, we adopt an adversarial training strategy to minimize the distribution distance between labeled and unlabeled data as inspired by domain adaptation works. On the other hand, to deal with the small sample size issue of labeled data, we also propose a simple interpolation strategy to generate pseudo training samples. Those two strategies can be easily implemented into existing deep neural networks. We demonstrate the effectiveness of our proposed approach on the benchmark SVHN and CIFAR10 datasets. Our code is available at https://github.com/qinenergy/adanet .

Person re-identification (ReID) aims at associating persons with the same identity across different views/scenes. Most existing methods improve matching accuracy by proposing high-dimensional real-valued features to represent person images comprehensively. However, considering the increasing data scale in real-world applications, the storage and matching efficiencies should be paid attention to as well. In this paper, we propose a binary coding approach for efficient ReID, inspired by the recent advances in adversarial learning. Specifically, the proposed Adversarial Binary Coding (ABC) implicitly fits the feature distribution to the expected binary one by optimizing the Wasserstein distance. To further enhance the semantic discriminability of binary codes, we seamlessly embed the ABC into a similarity measuring deep neural network. By end-to-end learning the framework, compact and discriminative binary features are generated for efficient and accurate ReID. Extensive experiments on large-scale benchmarks demonstrate the superiority of our approach over the state-of-the-art methods in both efficiency and accuracy.

Visual localization is a key step in many robotics pipelines, allowing the robot to (approximately) determine its position and orientation in the world. An efficient and scalable approach to visual localization is to use image retrieval techniques. These approaches identify the image most similar to a query photo in a database of geo-tagged images and approximate the query's pose via the pose of the retrieved database image. However, image retrieval across drastically different illumination conditions, e.g. day and night, is still a problem with unsatisfactory results, even in this age of powerful neural models. This is due to a lack of a suitably diverse dataset with true correspondences to perform end-to-end learning. A recent class of neural models allows for realistic translation of images among visual domains with relatively little training data and, most importantly, without ground-truth pairings.In this paper, we explore the task of accurately localizing images captured from two traversals of the same area in both day and night. We propose ToDayGAN - a modified image-translation model to alter nighttime driving images to a more useful daytime representation. We then compare the daytime and translated night images to obtain a pose estimate for the night image using the known 6-DOF position of the closest day image. Our approach improves localization performance by over 250% compared the current state-of-the-art, in the context of standard metrics in multiple categories.

Considerable progress has been made in semantic scene understanding of road scenes with monocular cameras. It is, however, mainly focused on certain specific classes such as cars, bicyclists and pedestrians. This work investigates traffic cones, an object category crucial for traffic control in the context of autonomous vehicles. 3D object detection using images from a monocular camera is intrinsically an ill-posed problem. In this work, we exploit the unique structure of traffic cones and propose a pipelined approach to solve this problem. Specifically, we first detect cones in images by a modified 2D object detector. Following which the keypoints on a traffic cone are recognized with the help of our deep structural regression network, here, the fact that the cross-ratio is projection invariant is leveraged for network regularization. Finally, the 3D position of cones is recovered via the classical Perspective n-Point algorithm using correspondences obtained from the keypoint regression. Extensive experiments show that our approach can accurately detect traffic cones and estimate their position in the 3D world in real time. The proposed method is also deployed on a real-time, autonomous system. It runs efficiently on the low-power Jetson TX2, providing accurate 3D position estimates, allowing a race-car to map and drive autonomously on an unseen track indicated by traffic cones. With the help of robust and accurate perception, our race-car won both Formula Student Competitions held in Italy and Germany in 2018, cruising at a top speed of 54 km/h on our driverless platform “gotthard driverless” https://youtu. be/HegmIXASKow?t==11694. Visualization of the complete pipeline, mapping and navigation can be found on our project page http://people.ee.ethz.ch/=tracczucrich/Trnfflcf.lonc/.

We present a general and flexible video-level framework for learning action models in videos. This method, called temporal segment network (TSN), aims to model long-range temporal structure with a new segment-based sampling and aggregation scheme. This unique design enables the TSN framework to efficiently learn action models by using the whole video. The learned models could be easily deployed for action recognition in both trimmed and untrimmed videos with simple average pooling and multi-scale temporal window integration, respectively. We also study a series of good practices for the implementation of the TSN framework given limited training samples. Our approach obtains the state-the-of-art performance on five challenging action recognition benchmarks: HMDB51 (71.0 percent), UCF101 (94.9 percent), THUMOS14 (80.1 percent), ActivityNet v1.2 (89.6 percent), and Kinetics400 (75.7 percent). In addition, using the proposed RGB difference as a simple motion representation, our method can still achieve competitive accuracy on UCF101 (91.0 percent) while running at 340 FPS. Furthermore, based on the proposed TSN framework, we won the video classification track at the ActivityNet challenge 2016 among 24 teams.

Video Object Segmentation, and video processing in general, has been historically dominated by methods that rely on the temporal consistency and redundancy in consecutive video frames. When the temporal smoothness is suddenly broken, such as when an object is occluded, or some frames are missing in a sequence, the result of these methods can deteriorate significantly. This paper explores the orthogonal approach of processing each frame independently, i.e., disregarding the temporal information. In particular, it tackles the task of semi-supervised video object segmentation: the separation of an object from the background in a video, given its mask in the first frame. We present Semantic One-Shot Video Object Segmentation (OSVOS$^\mathrm {S}$S), based on a fully-convolutional neural network architecture that is able to successively transfer generic semantic information, learned on ImageNet, to the task of foreground segmentation, and finally to learning the appearance of a single annotated object of the test sequence (hence one shot). We show that instance-level semantic information, when combined effectively, can dramatically improve the results of our previous method, OSVOS. We perform experiments on two recent single-object video segmentation databases, which show that OSVOS$^\mathrm {S}$S is both the fastest and most accurate method in the state of the art. Experiments on multi-object video segmentation show that OSVOS$^\mathrm {S}$S obtains competitive results.

Spatial understanding is a fundamental problem with wide-reaching real-world applications. The representation of spatial knowledge is often modeled with spatial templates, i.e., regions of acceptability of two objects under an explicit spatial relationship (e.g., "on," "below," etc.). In contrast with prior work that restricts spatial templates to explicit spatial prepositions (e.g., "glass on table"), here we extend this concept to implicit spatial language, i.e., those relationships (generally actions) for which the spatial arrangement of the objects is only implicitly implied (e.g., "man riding horse"). In contrast with explicit relationships, predicting spatial arrangements from implicit spatial language requires significant common sense spatial understanding. Here, we introduce the task of predicting spatial templates for two objects under a relationship, which can be seen as a spatial question-answering task with a (2D) continuous output ("where is the man w.r.t. a horse when the man is walking the horse?"). We present two simple neural-based models that leverage annotated images and structured text to learn this task. The good performance of these models reveals that spatial locations are to a large extent predictable from implicit spatial language. Crucially, the models attain similar performance in a challenging generalized setting, where the object-relation-object combinations (e.g., "man walking dog") have never been seen before. Next, we go one step further by presenting the models with unseen objects (e.g., "dog"). In this scenario, we show that leveraging word embeddings enables the models to output accurate spatial predictions, proving that the models acquire solid common sense spatial knowledge allowing for such generalization.

Learning representations on Grassmann manifolds is popular in quite a few visual recognition tasks. In order to enable deep learning on Grassmann manifolds, this paper proposes a deep network architecture by generalizing the Euclidean network paradigm to Grassmann manifolds. In particular, we design full rank mapping layers to transform input Grassmannian data to more desirable ones, exploit re-orthonormalization layers to normalize the resulting matrices, study projection pooling layers to reduce the model complexity in the Grassmannian context, and devise projection mapping layers to respect Grassmannian geometry and meanwhile achieve Euclidean forms for regular output layers. To train the Grassmann networks, we exploit a stochastic gradient descent setting on manifolds of the connection weights, and study a matrix generalization of backpropagation to update the structured data. The evaluations on three visual recognition tasks show that our Grassmann networks have clear advantages over existing Grassmann learning methods, and achieve results comparable with state-of-the-art approaches.

People detection methods are highly sensitive to occlusions between pedestrians, which are extremely frequent in many situations where cameras have to be mounted at a limited height. The reduction of camera prices allows for the generalization of static multi-camera set-ups. Using joint visual information from multiple synchronized cameras gives the opportunity to improve detection performance. In this paper, we present a new large-scale and high-resolution dataset. It has been captured with seven static cameras in a public open area, and unscripted dense groups of pedestrians standing and walking. Together with the camera frames, we provide an accurate joint (extrinsic and intrinsic) calibration, as well as 7 series of 400 annotated frames for detection at a rate of 2 frames per second. This results in over 40 000 bounding boxes delimiting every person present in the area of interest, for a total of more than 300 individuals. We provide a series of benchmark results using baseline algorithms published over the recent months for multi-view detection with deep neural networks, and trajectory estimation using a non-Markovian model.

Exploiting synthetic data to learn deep models has attracted increasing attention in recent years. However, the intrinsic domain difference between synthetic and real images usually causes a significant performance drop when applying the learned model to real world scenarios. This is mainly due to two reasons: 1) the model overfits to synthetic images, making the convolutional filters incompetent to extract informative representation for real images; 2) there is a distribution difference between synthetic and real data, which is also known as the domain adaptation problem. To this end, we propose a new reality oriented adaptation approach for urban scene semantic segmentation by learning from synthetic data. First, we propose a target guided distillation approach to learn the real image style, which is achieved by training the segmentation model to imitate a pretrained real style model using real images. Second, we further take advantage of the intrinsic spatial structure presented in urban scene images, and propose a spatial-aware adaptation scheme to effectively align the distribution of two domains. These two modules can be readily integrated with existing state-of-the-art semantic segmentation networks to improve their generalizability when adapting from synthetic to real urban scenes. We evaluate the proposed method on Cityscapes dataset by adapting from GTAV and SYNTHIA datasets, where the results demonstrate the effectiveness of our method.

Object detection typically assumes that training and test data are drawn from an identical distribution, which, however, does not always hold in practice. Such a distribution mismatch will lead to a significant performance drop. In this work, we aim to improve the cross-domain robustness of object detection. We tackle the domain shift on two levels: (1) the image-level shift, such as image style, illumination, etc., and (2) the instance-level shift, such as object appearance, size, etc. We build our approach based on the recent state-of-the-art Faster R-CNN model, and design two domain adaptation components, on image level and instance level, to reduce the domain discrepancy. The two domain adaptation components are based onH-divergence theory, and are implemented by learning a domain classifier in adversarial training manner. The domain classifiers on different levels are further reinforced with a consistency regularization to learn a domain-invariant region proposal network (RPN) in the Faster R-CNN model. We evaluate our newly proposed approach using multiple datasets including Cityscapes, KITTI, SIM10K, etc. The results demonstrate the effectiveness of our proposed approach for robust object detection in various domain shift scenarios.

This paper tackles the problem of video object segmentation, given some user annotation which indicates the object of interest. The problem is formulated as pixel-wise retrieval in a learned embedding space: we embed pixels of the same object instance into the vicinity of each other, using a fully convolutional network trained by a modified triplet loss as the embedding model. Then the annotated pixels are set as reference and the rest of the pixels are classified using a nearest-neighbor approach. The proposed method supports different kinds of user input such as segmentation mask in the first frame (semi-supervised scenario), or a sparse set of clicked points (interactive scenario). In the semi-supervised scenario, we achieve results competitive with the state of the art but at a fraction of computation cost (275 milliseconds per frame). In the interactive scenario where the user is able to refine their input iteratively, the proposed method provides instant response to each input, and reaches comparable quality to competing methods with much less interaction.

This paper introduces a novel variant of video summarization, namely building a summary that depends on the particular aspect of a video the viewer focuses on. We refer to this as viewpoint. To infer what the desired viewpoint may be, we assume that several other videos are available, especially groups of videos, e.g., as folders on a person's phone or laptop. The semantic similarity between videos in a group vs. the dissimilarity between groups is used to produce viewpoint-specific summaries. For considering similarity as well as avoiding redundancy, output summary should be (A) diverse, (B) representative of videos in the same group, and (C) discriminative against videos in the different groups. To satisfy these requirements (A)-(C) simultaneously, we proposed a novel video summarization method from multiple groups of videos. Inspired by Fisher's discriminant criteria, it selects summary by optimizing the combination of three terms (a) inner-summary, (b) inner-group, and (c) between-group variances defined on the feature representation of summary, which can simply represent (A)-(C). Moreover, we developed a novel dataset to investigate how well the generated summary reflects the underlying viewpoint. Quantitative and qualitative experiments conducted on the dataset demonstrate the effectiveness of proposed method.

Generating novel, yet realistic, images of persons is a challenging task due to the complex interplay between the different image factors, such as the foreground, background and pose information. In this work, we aim at generating such images based on a novel, two-stage reconstruction pipeline that learns a disentangled representation of the aforementioned image factors and generates novel person images at the same time. First, a multi-branched reconstruction network is proposed to disentangle and encode the three factors into embedding features, which are then combined to re-compose the input image itself. Second, three corresponding mapping functions are learned in an adversarial manner in order to map Gaussian noise to the learned embedding feature space, for each factor, respectively. Using the proposed framework, we can manipulate the foreground, background and pose of the input image, and also sample new embedding features to generate such targeted manipulations, that provide more control over the generation process. Experiments on the Market-1501 and Deepfashion datasets show that our model does not only generate realistic person images with new foregrounds, backgrounds and poses, but also manipulates the generated factors and interpolates the in-between states. Another set of experiments on Market-1501 shows that our model can also be beneficial for the person re-identification task1.

This paper explores the use of extreme points in an object (left-most, right-most, top, bottom pixels) as input to obtain precise object segmentation for images and videos. We do so by adding an extra channel to the image in the input of a convolutional neural network (CNN), which contains a Gaussian centered in each of the extreme points. The CNN learns to transform this information into a segmentation of an object that matches those extreme points. We demonstrate the usefulness of this approach for guided segmentation (grabcut-style), interactive segmentation, video object segmentation, and dense segmentation annotation. We show that we obtain the most precise results to date, also with less user input, in an extensive and varied selection of benchmarks and datasets. All our models and code are publicly available on http://www.vision.ee.ethz.ch/~cvlsegmentation/dextr/.

Deep Neural Networks trained as image auto-encoders have recently emerged as a promising direction for advancing the state-of-the-art in image compression. The key challenge in learning such networks is twofold: To deal with quantization, and to control the trade-off between reconstruction error (distortion) and entropy (rate) of the latent image representation. In this paper, we focus on the latter challenge and propose a new technique to navigate the rate-distortion trade-off for an image compression auto-encoder. The main idea is to directly model the entropy of the latent representation by using a context model: A 3D-CNN which learns a conditional probability model of the latent distribution of the auto-encoder. During training, the auto-encoder makes use of the context model to estimate the entropy of its representation, and the context model is concurrently updated to learn the dependencies between the symbols in the latent representation. Our experiments show that this approach, when measured in MS-SSIM, yields a state-of-the-art image compression system based on a simple convolutional auto-encoder.

In this paper, we consider the problem of reconstructing a dense 3D model using images captured from different views. Recent methods based on convolutional neural networks (CNN) allow learning the entire task from data. However, they do not incorporate the physics of image formation such as perspective geometry and occlusion. Instead, classical approaches based on Markov Random Fields (MRF) with ray-potentials explicitly model these physical processes, but they cannot cope with large surface appearance variations across different viewpoints. In this paper, we propose RayNet, which combines the strengths of both frameworks. RayNet integrates a CNN that learns view-invariant feature representations with an MRF that explicitly encodes the physics of perspective projection and occlusion. We train RayNet end-to-end using empirical risk minimization. We thoroughly evaluate our approach on challenging real-world datasets and demonstrate its benefits over a piece-wise trained baseline, hand-crafted models as well as other learning-based approaches.

Designing a logo for a new brand is a lengthy and tedious back-and-forth process between a designer and a client. In this paper we explore to what extent machine learning can solve the creative task of the designer. For this, we build a dataset - LLD - of 600k+ logos crawled from the world wide web. Training Generative Adversarial Networks (GANs) for logo synthesis on such multi-modal data is not straightforward and results in mode collapse for some state-of-the-art methods. We propose the use of synthetic labels obtained through clustering to disentangle and stabilize GAN training, and validate this approach on CIFAR-10 and ImageNet-small to demonstrate its generality. We are able to generate a high diversity of plausible logos and demonstrate latent space exploration techniques to ease the logo design task in an interactive manner. GANs can cope with multi-modal data by means of synthetic labels achieved through clustering, and our results show the creative potential of such techniques for logo synthesis and manipulation. Our dataset and models are publicly available at https://data.vision.ee.ethz.ch/sagea/lld/.

Convolutional neural networks rely on image texture and structure to serve as discriminative features to classify the image content. Image enhancement techniques can be used as preprocessing steps to help improve the overall image quality and in turn improve the overall effectiveness of a CNN. Existing image enhancement methods, however, are designed to improve the perceptual quality of an image for a human observer. In this paper, we are interested in learning CNNs that can emulate image enhancement and restoration, but with the overall goal to improve image classification and not necessarily human perception. To this end, we present a unified CNN architecture that uses a range of enhancement filters that can enhance image-specific details via end-to-end dynamic filter learning. We demonstrate the effectiveness of this strategy on four challenging benchmark datasets for fine-grained, object, scene, and texture classification: CUB-200-2011, PASCAL-VOC2007, MIT-Indoor, and DTD. Experiments using our proposed enhancement show promising results on all the datasets. In addition, our approach is capable of improving the performance of all generic CNN architectures.

We propose a novel method for the geometric registration of semantically labeled regions. We approximate semantic regions by ellipsoids, and leverage their convexity to formulate the correspondence search effectively as a constrained optimization problem that maximizes the number of matched regions, and which we solve globally optimal in a Branch-and-Bound fashion. To this end, we derive suitable linear matrix inequality constraints which describe ellipsoid-to-ellipsoid assignment conditions. Our approach is robust to large percentages of outliers and thus applicable to difficult correspondence search problems. In multiple experiments we demonstrate the flexibility and robustness of our approach on a number of challenging vision problems.

As more and more personal photos are shared online, being able to obfuscate identities in such photos is becoming a necessity for privacy protection. People have largely resorted to blacking out or blurring head regions, but they result in poor user experience while being surprisingly ineffective against state of the art person recognizers [17]. In this work, we propose a novel head inpainting obfuscation technique. Generating a realistic head inpainting in social media photos is challenging because subjects appear in diverse activities and head orientations. We thus split the task into two sub-tasks: (1) facial landmark generation from image context (e.g. body pose) for seamless hypothesis of sensible head pose, and (2) facial landmark conditioned head inpainting. We verify that our inpainting method generates realistic person images, while achieving superior obfuscation performance against automatic person recognizers.

We investigate the problem of object referring (OR) i.e. to localize a target object in a visual scene coming with a language description. Humans perceive the world more as continued video snippets than as static images, and describe objects not only by their appearance, but also by their spatio-temporal context and motion features. Humans also gaze at the object when they issue a referring expression. Existing works for OR mostly focus on static images only, which fall short in providing many such cues. This paper addresses OR in videos with language and human gaze. To that end, we present a new video dataset for OR, with 30, 000 objects over 5, 000 stereo video sequences annotated for their descriptions and gaze. We further propose a novel network model for OR in videos, by integrating appearance, motion, gaze, and spatio-temporal context into one network. Experimental results show that our method effectively utilizes motion cues, human gaze, and spatio-temporal context. Our method outperforms previous OR methods. For dataset and code, please refer https://people.ee.ethz.ch/~arunv/ORGaze.html.

We present a simple and effective method for 3D hand pose estimation from a single depth frame. As opposed to previous state-of-the-art methods based on holistic 3D regression, our method works on dense pixel-wise estimation. This is achieved by careful design choices in pose parameterization, which leverages both 2D and 3D properties of depth map. Specifically, we decompose the pose parameters into a set of per-pixel estimations, i.e., 2D heat maps, 3D heat maps and unit 3D directional vector fields. The 2D/3D joint heat maps and 3D joint offsets are estimated via multitask network cascades, which is trained end-to-end. The pixel-wise estimations can be directly translated into a vote casting scheme. A variant of mean shift is then used to aggregate local votes while enforcing consensus between the the estimated 3D pose and the pixel-wise 2D and 3D estimations by design. Our method is efficient and highly accurate. On MSRA and NYU hand dataset, our method outperforms all previous state-of-the-art approaches by a large margin. On the ICVL hand dataset, our method achieves similar accuracy compared to the nearly saturated result obtained by [5] and outperforms various other proposed methods. Code is available online1.

Spatiotemporal feature learning in videos is a fundamental problem in computer vision. This paper presents a new architecture, termed as Appearance-and-Relation Network (ARTNet), to learn video representation in an end-to-end manner. ARTNets are constructed by stacking multiple generic building blocks, called as SMART, whose goal is to simultaneously model appearance and relation from RGB input in a separate and explicit manner. Specifically, SMART blocks decouple the spatiotemporal learning module into an appearance branch for spatial modeling and a relation branch for temporal modeling. The appearance branch is implemented based on the linear combination of pixels or filter responses in each frame, while the relation branch is designed based on the multiplicative interactions between pixels or filter responses across multiple frames. We perform experiments on three action recognition benchmarks: Kinetics, UCF101, and HMDB51, demonstrating that SMART blocks obtain an evident improvement over 3D convolutions for spatiotemporal feature learning. Under the same training setting, ARTNets achieve superior performance on these three datasets to the existing state-of-the-art methods.1