Recently, image-to-video (I2V) diffusion models have demonstrated impressive scene understanding and generative quality, incorporating image conditions to guide generation. However, these models primarily animate static images without extending beyond their provided context. Introducing additional constraints, such as camera trajectories, can enhance diversity but often degrade visual quality, limiting their applicability for tasks requiring faithful scene representation. We propose CamC2V, a context-to-video$(C 2 V)$model that integrates multiple image conditions as context with 3D constraints alongside camera control to enrich both global semantics and fine-grained visual details. This enables more coherent and context-aware video generation. Moreover, we motivate the necessity of temporal awareness for an effective context representation. Our comprehensive study on the RealEstate10K dataset demonstrates a 24.09 % (FVD) improvement in visual quality and camera controllability. Our code is publicly available at: https://github.com/LDenninger/CamC2V.

Recent advances in action anticipation rely heavily on Transformer architectures to learn discriminative representations of the past observation, incurring high computational and memory overhead that limits their applicability to long videos. While temporal processors with linear complexity like RNNs and state-space models offer efficient alternatives, their sequential nature risks overlooking subtle cues in observed frames that could enhance future anticipation. We address this limitation with Cross Linear Attentive Memory (CLAM), a memory module that selectively retrieves complementary context cues from frame features. By reformulating linear attention to replace traditional cross-attention, CLAM achieves linear computation complexity and constant memory usage relative to input length. Finally, by fusing the outputs of the temporal processor and CLAM, a non-autoregressive Transformer decoder generates future actions in one shot with high accuracy. Experiments on egocentric (EpicKitchens100 and Ego4D) and third-person (Thumos14) benchmarks demonstrate our model’s superior anticipation accuracy and scalability, processing longer sequences with significantly less latency growth than alternatives. Our approach also achieves promising results in online action detection.

We introduce Action Discovery, a novel task that addresses the challenge of discovering actions in long, untrimmed videos where only a subset of the present actions have been annotated. The goalis thus to discover new actions in the video segments that have not been annotated or annotated by a generic background class. This scenario is particularly relevant in domains like neuroscience, where well-defined behaviors (e.g., walking, eating) coexist with subtle or infrequent actions that are often overlooked, as well as in applications where datasets are inherently partially annotateddue to ambiguous or missing labels. To address this problem, we propose a two-step approach that leverages the known annotations to guide both the temporal and semantic granularity of unknownaction segments. First, we introduce the Granularity-Guided Segmentation Module (GGSM), which identifies temporal intervals for both known and unknown actions by mimicking the granularity ofannotated actions. Second, we propose the Unknown Action Segment Assignment (UASA), which identifies semantically meaningful classes within the unknown actions, based on learned embedding similarities. We systematically explore the proposed setting of Action Discovery on three challenging datasets - Breakfast, 50Salads, and Desktop Assembly - demonstrating that our method considerablyimproves upon existing baselines.

Abstract While there has been substantial progress in temporal action segmentation, the challenge to generalize to unseen views remains unaddressed. Hence, we define a protocol for unseen view action segmentation where camera views for evaluating the model are unavailable during training. This includes changing from top-frontal views to a side view or even more challenging from exocentric to egocentric views. Furthermore, we present an approach for temporal action segmentation that tackles this challenge. Our approach leverages a shared representation at both the sequence and segment levels to reduce the impact of view differences during training. We achieve this by introducing a sequence loss and an action loss, which together facilitate consistent video and action representations across different views. The evaluation on the Assembly101, IkeaASM, and EgoExoLearn datasets demonstrate significant improvements, with a $$12.8\%$$ 12.8 % increase in F1@50 for unseen exocentric views and a substantial $$54\%$$ 54 % improvement for unseen egocentric views.

Abstract State-space models (SSMs), exemplified by S4, have introduced a novel context modeling method by integrating state-space techniques into deep learning. Despite their effectiveness, SSMs struggle with global context modeling due to data-independent matrices. The Mamba model addresses this with data-dependent variants enabled by the S6 selective-scan algorithm, enhancing context modeling, especially for long sequences. However, Mamba-based architectures face significant parameter scalability challenges, limiting their utility in vision applications. This paper tackles the scalability issue of large SSMs for image classification and action recognition without relying on additional techniques like knowledge distillation. We analyze the distinct characteristics of Mamba-based and Attention-based models, proposing a Mamba-Attention interleaved architecture that enhances scalability, robustness, and performance. We demonstrate that the stable and efficient interleaved architecture resolves the scalability issue of Mamba-based architectures and increases robustness to common corruption artifacts. Our thorough evaluation on the ImageNet-1K, Kinetics-400, and Something-Something-v2 benchmarks demonstrates that our approach improves the accuracy of state-of-the-art Mamba-based architectures by up to $$+1.7$$ + 1.7 %.

Many query-based approaches for 3D Multi-Object Tracking (MOT) adopt the tracking-by-attention paradigm, utilizing track queries for identity-consistent detection and object queries for identity-agnostic track spawning. Tracking-by-attention, however, entangles detection and tracking queries in one embedding for both the detection and tracking task, which is sub-optimal. Other approaches resemble the tracking-by-detection paradigm and detect objects using decoupled track and detection queries followed by a subsequent association. These methods, however, do not leverage synergies between the detection and association task. Combining the strengths of both paradigms, we introduce ADA-Track++, a novel end-to-end framework for 3D MOT from multi-view cameras. We introduce a learnable data association module based on edge-augmented cross-attention, leveraging appearance and geometric features. We also propose an auxiliary token in this attention-based association module, which helps mitigate disproportionately high attention to incorrect association targets caused by attention normalization. Furthermore, we integrate this association module into the decoder layer of a DETR-based 3D detector, enabling simultaneous DETR-like query-to-image cross-attention for detection and query-to-query cross-attention for data association. By stacking these decoder layers, queries are refined for the detection and association task alternately, effectively harnessing the task dependencies. We evaluate our method on the nuScenes dataset and demonstrate the advantage of our approach compared to the two previous paradigms.

In this work, we address unsupervised temporal action segmentation, which segments a set of long, untrimmed videos into semantically meaningful segments that are consistent across videos. While recent approaches combine representation learning and clustering in a single step for this task, they do not cope with large variations within temporal segments of the same class. To address this limitation, we propose a novel method, termed Hierarchical Vector Quantization (HVQ), that consists of two subsequent vector quantization modules. This results in a hierarchical clustering where the additional subclusters cover the variations within a cluster. We demonstrate that our approach captures the distribution of segment lengths much better than the state of the art. To this end, we introduce a new metric based on the Jensen-Shannon Distance (JSD) for unsupervised temporal action segmentation. We evaluate our approach on three public datasets, namely Breakfast, YouTube Instructional and IKEA ASM. Our approach outperforms the state of the art in terms of F1 score, recall and JSD.

Predicting future video frames is essential for decision-making systems, yet RGB frames alone often lack the information needed to fully capture the underlying complexities of the real world. To address this limitation, we propose a multi-modal framework for Synchronous Video Prediction (SyncVP) that incorporates complementary data modalities, enhancing the richness and accuracy of future predictions. SyncVP builds on pre-trained modality-specific diffusion models and introduces an efficient spatio-temporal cross-attention module to enable effective information sharing across modalities. We evaluate SyncVP on standard benchmark datasets, such as Cityscapes and BAIR, using depth as an additional modality. We furthermore demonstrate its generalization to other modalities on SYNTHIA with semantic information and ERA5-Land with climate data. Notably, SyncVP achieves state-of-the-art performance, even in scenarios where only one modality is present, demonstrating its robustness and potential for a wide range of applications.

State-Space models (SSMs) have recently shown promise in capturing long-range dependencies with subquadratic computational complexity, making them attractive for various applications. However, purely SSM-based models face critical challenges related to stability and achieving state-of-the-art performance in computer vision tasks. Our paper addresses the challenges of scaling SSM-based models for computer vision, particularly the instability and inefficiency of large model sizes. We introduce a parameter-efficient modulated group mamba layer that divides the input channels into four groups and applies our proposed SSM-based efficient Visual Single Selective Scanning (VSSS) block independently to each group, with each VSSS block scanning in one of the four spatial directions. The Modulated Group Mamba layer also wraps the four VSSS blocks into a channel modulation operator to improve cross-channel communication. Furthermore, we introduce a distillation-based training objective to stabilize the training of large models, leading to consistent performance gains. Our comprehensive experiments demonstrate the merits of the proposed contributions, leading to superior performance over existing methods for image classification on ImageNet-1K, object detection, instance segmentation on MS-COCO, and semantic segmentation on ADE20K. Our tiny variant with 23M parameters achieves state-of-the-art performance with a classification top-1 accuracy of 83.3% on ImageNet-1K, while being 26% efficient in terms of parameters, compared to the best existing Mamba design of same model size. Code and models are available at: https://github.com/Amshaker/GroupMamba

Advancements in Computer-Aided Screening (CAS) systems are essential for improving the detection of security threats in X-ray baggage scans. However, current datasets are limited in representing real-world, sophisticated threats and concealment tactics, and existing approaches are constrained by a closed-set paradigm with predefined labels. To address these challenges, we introduce STCray, the first multimodal X-ray baggage security dataset, comprising 46,642 image-caption paired scans across 21 threat categories, generated using an X-ray scanner for airport security. STCray is meticulously developed with our specialized protocol that ensures domain-aware, coherent captions, that lead to the multi-modal instruction following data in X-ray baggage security. This allows us to train a domain-aware visual AI assistant named STING-BEE that supports a range of vision-language tasks, including scene comprehension, referring threat localization, visual grounding, and visual question answering (VQA), establishing novel baselines for multi-modal learning in X-ray baggage security. Further, STING-BEE shows state-of-the-art generalization in cross-domain settings. Code, data, and models are available at https://divs1159.github.io/STING-BEE/.

We present an efficient encoder-free approach for video-language understanding that achieves competitive performance while significantly reducing computational overhead. Current video-language models typically rely on heavyweight image encoders (300M-1.1B parameters) or video encoders (1B-1.4B parameters), creating a substantial computational burden when processing multi-frame videos. Our method introduces a novel Spatio-Temporal Alignment Block (STAB) that directly processes video inputs without requiring pre-trained encoders while using only 45M parameters for visual processing - at least a 6.5× reduction compared to traditional approaches. The STAB architecture combines Local Spatio-Temporal Encoding for fine-grained feature extraction, efficient spatial downsampling through learned attention and separate mechanisms for modeling frame-level and video-level relationships. Our model achieves comparable or superior performance to encoder-based approaches for open-ended video question answering on standard benchmarks. The fine-grained video question-answering evaluation demonstrates our model’s effectiveness, outperforming the encoder-based approaches Video-ChatGPT and Video-LLaVA in key aspects like correctness and temporal understanding. Extensive ablation studies validate our architectural choices and demonstrate the effectiveness of our spatio-temporal modeling approach while achieving 3-4× faster processing speeds than previous methods. Code is available at https://jh-yi.github.io/Video-Panda.

Long-term dense action anticipation is very challenging since it requires predicting actions and their durations several minutes into the future based on provided video observations. To model the uncertainty of future outcomes, stochastic models predict several potential future action sequences for the same observation. Recent work has further proposed to incorporate uncertainty modelling for observed frames by simultaneously predicting per-frame past and future actions in a unified manner. While such joint modelling of actions is beneficial, it requires long-range temporal capabilities to connect events across distant past and future time points. However, the previous work struggles to achieve such a long-range understanding due to its limited and/or sparse receptive field. To alleviate this issue, we propose a novel MANTA (MAmbafor ANTicipation) network. Our model enables effective long-term temporal modelling even for very long sequences while maintaining linear complexity in sequence length. We demonstrate that our approach achieves state-of-the-art results on three datasets—Breakfast, 50Salads, and Assembly 101—while also significantly improving computational and memory efficiency. Our code is available at https://github.com/olga-zats/DlFFMANTA.

Video-based Person Re-identification (ReID) is crucial in visual surveillance, focusing on matching video snippets of individuals across multiple non-overlapping cameras. Existing methods either conduct ReID at the image level without leveraging temporal information, or employ complex temporal information aggregation techniques, which results in substantial network size and reduced performance efficiency. Recent advances in Vision Transformer (ViT) architectures leverage diverse large-scale datasets alongside sophisticated architectures to achieve enhanced fine-grained feature discrimination. To fully explore the potential of ViT architectures without adding substantial additional modules for video-based ReID, we propose PeT-KeyStAtion: a Parameter-efficient Transformer with Keypoint- guided Spatial-temporal Aggregation using a Spatial-Temporal and Keypoint (STK) Module with lightweight adapters. Our framework effectively captures and aggregates spatial, temporal, and keypoint information with only 11% of the parameters compared to full fine-tuning. Extensive experiments show that our method outperforms state-of-the-art baselines on MARS and iLIDS-VID, and achieves promising performance on LS-VID.

We present MixANT, a novel architecture for stochastic long-term dense anticipation of human activities. While recent State Space Models (SSMs) like Mamba have shown promise through input-dependent selectivity on three key parameters, the critical forget-gate ($\textbf{A}$ matrix) controlling temporal memory remains static. We address this limitation by introducing a mixture of experts approach that dynamically selects contextually relevant $\textbf{A}$ matrices based on input features, enhancing representational capacity without sacrificing computational efficiency. Extensive experiments on the 50Salads, Breakfast, and Assembly101 datasets demonstrate that MixANT consistently outperforms state-of-the-art methods across all evaluation settings. Our results highlight the importance of input-dependent forget-gate mechanisms for reliable prediction of human behavior in diverse real-world scenarios.

Modern metrics for generative learning like Fréchet Inception Distance (FID) and DINOv2-Fréchet Distance (FD-DINOv2) demonstrate impressive performance. However, they suffer from various shortcomings, like a bias towards specific generators and datasets. To address this problem, we propose the Fréchet Wavelet Distance (FWD) as a domain-agnostic metric based on the Wavelet Packet Transform ($\mathcal{W}_p$). FWD provides a sight across a broad spectrum of frequencies in images with a high resolution, preserving both spatial and textural aspects. Specifically, we use $\mathcal{W}_p$ to project generated and real images to the packet coefficient space. We then compute the Fréchet distance with the resultant coefficients to evaluate the quality of a generator. This metric is general-purpose and dataset-domain agnostic, as it does not rely on any pre-trained network, while being more interpretable due to its ability to compute Fréchet distance per packet, enhancing transparency. We conclude with an extensive evaluation of a wide variety of generators across various datasets that the proposed FWD can generalize and improve robustness to domain shifts and various corruptions compared to other metrics.

Modern methods for fine-tuning a Vision Transformer (ViT) like Low-Rank Adaptation (LoRA) and its variants demonstrate impressive performance. However, these methods ignore the high-dimensional nature of Multi-Head Attention (MHA) weight tensors. To address this limitation, we propose Canonical Rank Adaptation (CaRA). CaRA leverages tensor mathematics, first by tensorising the transformer into two different tensors; one for projection layers in MHA and the other for feed-forward layers. Second, the tensorised formulation is fine-tuned using the low-rank adaptation in Canonical-Polyadic Decomposition (CPD) form. Employing CaRA efficiently minimizes the number of trainable parameters. Experimentally, CaRA outperforms existing Parameter-Efficient Fine-Tuning (PEFT) methods in visual classification benchmarks such as Visual Task Adaptation Benchmark (VTAB)-1k and Fine-Grained Visual Categorization (FGVC).

Recent deep learning approaches for river discharge forecasting have improved the accuracy and efficiency in flood forecasting, enabling more reliable early warning systems for risk management. Nevertheless, existing deep learning approaches in hydrology remain largely confined to local-scale applications and do not leverage the inherent spatial connections of bodies of water. Thus, there is a strong need for new deep learning methodologies that are capable of modeling spatio-temporal relations to improve river discharge and flood forecasting for scientific and operational applications. To address this, we present RiverMamba, a novel deep learning model that is pretrained with long-term reanalysis data and that can forecast global river discharge and floods on a $0.05^\circ$ grid up to 7 days lead time, which is of high relevance in early warning. To achieve this, RiverMamba leverages efficient Mamba blocks that enable the model to capture spatio-temporal relations in very large river networks and enhance its forecast capability for longer lead times. The forecast blocks integrate ECMWF HRES meteorological forecasts, while accounting for their inaccuracies through spatio-temporal modeling. Our analysis demonstrates that RiverMamba provides reliable predictions of river discharge across various flood return periods, including extreme floods, and lead times, surpassing both AI- and physics-based models. The source code and datasets are publicly available at the project page https://hakamshams.github.io/RiverMamba.

Novel View Synthesis (NVS), which tries to produce a realistic image at the target view given source view images and their corresponding poses, is a fundamental problem in 3D Vision. As this task is heavily under-constrained, some recent work, like Zerol23 [18], tries to solve this problem with generative modeling, specifically using pre-trained diffusion models. Although this strategy generalizes well to new scenes, compared to neural radiance field-based methods, it offers low levels of flexibility. For example, it can only accept a single-view image as input, despite realistic applications often offering multiple input images. This is because the source-view images and corresponding poses are processed separately and injected into the model at different stages. Thus it is not trivial to generalize the model into multi-view source images, once they are available. To solve this issue, we try to process each pose image pair separately and then fuse them as a unified visual representation which will be injected into the model to guide image synthesis at the target-views. However, inconsistency and computation costs increase as the number of input source-view images increases. To solve these issues, the Multi-view Cross Former module is proposed which maps variable-length input data to fix-size output data. A two-stage training strategy is introduced to further improve the efficiency during training time. Qualitative and quantitative evaluation over multiple datasets demonstrates the effectiveness of the proposed method against previous approaches. The code will be released according to the acceptance.

Many point-based semantic segmentation methods have been designed for indoor scenarios, but they struggle if they are applied to point clouds that are captured by a light detection and ranging (LiDAR) sensor in an outdoor environment. In order to make these methods more efficient and robust such that they can handle LiDAR data, we introduce the general concept of reformulating 3-D point-based operations such that they can operate in the projection space. While we show by means of three point-based methods that the reformulated versions are between 300 and 400 times faster and achieve higher accuracy, we furthermore demonstrate that the concept of reformulating 3-D point-based operations allows to design new architectures that unify the benefits of point-based and image-based methods. As an example, we introduce a network that integrates reformulated 3-D point-based operations into a 2-D encoder-decoder architecture that fuses the information from different 2-D scales. We evaluate the approach on four challenging datasets for semantic LiDAR point cloud segmentation and show that leveraging reformulated 3-D point-based operations with 2-D image-based operations achieves very good results for all four datasets.

Many query-based approaches for 3D Multi-Object Tracking (MOT) adopt the tracking-by-attention paradigm, utilizing track queries for identity-consistent detection and object queries for identity-agnostic track spawning. Tracking-by-attention, however, entangles detection and tracking queries in one embedding for both the detection and tracking task, which is sub-optimal. Other approaches resemble the tracking-by-detection paradigm, detecting objects using decoupled track and detection queries followed by a subsequent association. These methods, however, do not leverage synergies between the detection and association task. Combining the strengths of both paradigms, we introduce ADA-Track, a novel end-to-endframeworkfor 3D MOT from multi-view cameras. We introduce a learnable data association module based on edge-augmented cross-attention, leveraging appearance and geometric features. Furthermore, we integrate this association module into the decoder layer of a DETR-based 3D detector, enabling simultaneous DETR-like query-to-image cross-attention for detection and query-to-query cross-attention for data association. By stacking these decoder layers, queries are refined for the detection and association task alternately, effectively harnessing the task dependencies. We evaluate our method on the nuScenes dataset and demonstrate the advantage of our approach compared to the two previous paradigms. Code is available at https: / / gi thub. com/ dsx0511/ADA- Track.

Counting repetitive actions in long untrimmed videos is a challenging task that has many applications such as rehabilitation. State-of-the-art methods predict action counts by first generating a temporal self-similarity matrix (TSM) from the sampled frames and then feeding the matrix to a predictor network. The self-similarity matrix, however, is not an optimal input to a network since it discards too much information from the frame-wise embeddings. We thus rethink how a TSM can be utilized for counting repetitive actions and propose a framework that learns embeddings and predicts action start probabilities at full temporal resolution. The number of repeated actions is then inferred from the action start probabilities. In contrast to current approaches that have the TSM as an intermediate representation, we propose a novel loss based on a generated reference TSM, which enforces that the self-similarity of the learned frame-wise embeddings is consistent with the self-similarity of repeated actions. The proposed framework achieves state-of-the-art results on three datasets, i.e., RepCount, UCFRep, and Countix.

An object handover between a robot and a human is a coordinated action which is prone to failure for reasons such as miscommunication, incorrect actions and unexpected object properties. Existing works on handover failure detection and prevention focus on preventing failures due to object slip or external disturbances. However, there is a lack of datasets and evaluation methods that consider unpreventable failures caused by the human participant. To address this deficit, we present the multimodal Handover Failure Detection dataset, which consists of failures induced by the human participant, such as ignoring the robot or not releasing the object. We also present two baseline methods for handover failure detection: (i) a video classification method using 3D CNNs and (ii) a temporal action segmentation approach which jointly classifies the human action, robot action and overall outcome of the action. The results show that video is an important modality, but using force-torque data and gripper position help improve failure detection and action segmentation accuracy.

The spatio-temporal relations of impacts of extreme events and their drivers in climate data are not fully understood and there is a need of machine learning approaches to identify such spatio-temporal relations from data. The task, however, is very challenging since there are time delays between extremes and their drivers, and the spatial response of such drivers is inhomogeneous. In this work, we propose a first approach and benchmarks to tackle this challenge. Our approach is trained end-to-end to predict spatio-temporally extremes and spatio-temporally drivers in the physical input variables jointly. By enforcing the network to predict extremes from spatio-temporal binary masks of identified drivers, the network successfully identifies drivers that are correlated with extremes. We evaluate our approach on three newly created synthetic benchmarks, where two of them are based on remote sensing or reanalysis climate data, and on two real-world reanalysis datasets. The source code and datasets are publicly available at the project page https://hakamshams.github.io/IDE.

Modeling long-term context in videos is crucial for many fine-grained tasks including temporal action segmentation. An interesting question that is still open is how much long-term temporal context is needed for optimal performance. While transformers can model the long-term context of a video, this becomes computationally prohibitive for long videos. Recent works on temporal action segmentation thus combine temporal convolutional networks with self-attentions that are computed only for a local temporal window. While these approaches show good results, their performance is limited by their inability to capture the full context of a video. In this work, we try to answer how much long-term temporal context is required for temporal action segmentation by introducing a transformer-based model that leverages sparse attention to capture the full context of a video. We compare our model with the current state of the art on three datasets for temporal action segmentation, namely 50Salads, Breakfast, and Assembly101. Our experiments show that modeling the full context of a video is necessary to obtain the best performance for temporal action segmentation.

Tracking 3D objects accurately and consistently is crucial for autonomous vehicles, enabling more reliable downstream tasks such as trajectory prediction and motion planning. Based on the substantial progress in object detection in recent years, the tracking-by-detection paradigm has become a popular choice due to its simplicity and efficiency. State-of-the-art 3D multi-object tracking (MOT) approaches typically rely on non-learned model-based algorithms such as Kalman Filter but require many manually tuned parameters. On the other hand, learning-based approaches face the problem of adapting the training to the online setting, leading to inevitable distribution mismatch between training and inference as well as suboptimal performance. In this work, we propose 3DMOTFormer, a learned geometry-based 3D MOT framework building upon the transformer architecture. We use an Edge-Augmented Graph Transformer to reason on the track-detection bipartite graph frame-by-frame and conduct data association via edge classification. To reduce the distribution mismatch between training and inference, we propose a novel online training strategy with an autoregressive and recurrent forward pass as well as sequential batch optimization. Using CenterPoint detections, our approach achieves 71.2% and 68.2% AMOTA on the nuScenes validation and test split, respectively. In addition, a trained 3DMOTFormer model generalizes well across different object detectors. Code is available at: https://github.com/dsx0511/3DMOTFormer.

The task of few-shot GAN adaptation aims to adapt a pre-trained GAN model to a small dataset with very few training images. While existing methods perform well when the dataset for pre-training is structurally similar to the target dataset, the approaches suffer from training instabilities or memorization issues when the objects in the two domains have a very different structure. To mitigate this limitation, we propose a new smoothness similarity regularization that transfers the inherently learned smoothness of the pre-trained GAN to the few-shot target domain even if the two domains are very different. We evaluate our approach by adapting an unconditional and a class-conditional GAN to diverse few-shot target domains. Our proposed method significantly outperforms prior few-shot GAN adaptation methods in the challenging case of structurally dissimilar source-target domains, while performing on par with the state of the art for similar source-target domains.

We propose Social Diffusion, a novel method for short-term and long-term forecasting of the motion of multiple persons as well as their social interactions. Jointly forecasting motions for multiple persons involved in social activities is inherently a challenging problem due to the interdependencies between individuals. In this work, we leverage a diffusion model conditioned on motion histories and causal temporal convolutional networks to forecast individually and contextually plausible motions for all participants. The contextual plausibility is achieved via an order-invariant aggregation function. As a second contribution, we design a new evaluation protocol that measures the plausibility of social interactions which we evaluate on the Haggling dataset, which features a challenging social activity where people are actively taking turns to talk and switching their attention. We evaluate our approach on four datasets for multi-person forecasting where our approach outperforms the state-of-the-art in terms of motion realism and contextual plausibility.

In this paper, we address the problem of short-term action anticipation, i.e., we want to predict an upcoming action one second before it happens. We propose to harness high-level intent information to anticipate actions that will take place in the future. To this end, we incorporate an additional goal prediction branch into our model and propose a consistency loss function that encourages the anticipated actions to conform to the high-level goal pursued in the video. In our experiments, we show the effectiveness of the proposed approach and demonstrate that our method achieves state-of-the-art results on two large-scale datasets: Assembly101 and COIN. The code is available at https://github.com/olga-zats/goal_consistency.

Remote sensing has enabled large-scale crop classification to understand agricultural ecosystems and estimate production yields. Since few years, machine learning is increasingly used for automated crop classification. However, most approaches apply novel algorithms to custom datasets containing information of few crop fields covering a small region and this often leads to poor models that lack generalization capability. Therefore in this work, inspired from the self-supervised learning approaches, we devised and compared different approaches for contrastive self-supervised learning using Sentinel2 and Planetscope data for crop classification. In addition, based on the dataset DENETHOR, we assembled our own dataset for the experiments.

Accurately perceiving instances and predicting their future motion are key tasks for autonomous vehicles, enabling them to navigate safely in complex urban traffic. While bird’s-eye view (BEV) representations are commonplace in perception for autonomous driving, their potential in a motion prediction setting is less explored. Existing approaches for BEV instance prediction from surround cameras rely on a multi-task auto-regressive setup coupled with complex post-processing to predict future instances in a spatio-temporally consistent manner. In this paper, we depart from this paradigm and propose an efficient novel end-to-end framework named PowerBEV, which differs in several design choices aimed at reducing the inherent redundancy in previous methods. First, rather than predicting the future in an auto-regressive fashion, PowerBEV uses a parallel, multi-scale module built from lightweight 2D convolutional networks. Second, we show that segmentation and centripetal backward flow are sufficient for prediction, simplifying previous multi-task objectives by eliminating redundant output modalities. Building on this output representation, we propose a simple, flow warping-based post-processing approach which produces more stable instance associations across time. Through this lightweight yet powerful design, PowerBEV outperforms state-of-the-art baselines on the NuScenes Dataset and poses an alternative paradigm for BEV instance prediction. We made our code publicly available at: https://github.com/EdwardLeeLPZ/PowerBEV.

In this talk, I will first discuss approaches that reduce the GFLOPs during inference for 3D convolutional neural networks (CNN) and vision transformers. While state-of-the-art 3D CNNs and vision transformers achieve very good results on action recognition datasets, they are computationally very expensive and require many GFLOPs. While the GFLOPs of a 3D CNN or vision transformer can be decreased by reducing the temporal feature resolution or the number of tokens, there is no setting that is optimal for all input clips. I will therefore discuss two differentiable sampling approaches that can be plugged into any existing 3D CNN or vision transformer architecture. The sampling approaches adapt the computational resources to the input video such that as much resources as needed but not more than necessary are used to classify a video. The approaches substantially reduce the computational cost (GFLOPs) of state-of-the-art networks while preserving the accuracy. In the second part, I will discuss an approach that generates annotated training samples of very rare classes. It is based on a generative adversarial network (GAN) that jointly synthesizes images and the corresponding segmentation mask for each image. The generated data can then be used for one-shot video object segmentation.

Forecasting human motion of multiple persons is very challenging. It requires to model the interactions between humans and the interactions with objects and the environment. For example, a person might want to make a coffee, but if the coffee machine is already occupied the person will haveto wait. These complex relations between scene geometry and persons ariseconstantly in our daily lives, and models that wish to accurately forecasthuman behavior will have to take them into consideration. To facilitate research in this direction, we propose Humans in Kitchens, alarge-scale multi-person human motion dataset with annotated 3D human poses, scene geometry and activities per person and frame.Our dataset consists of over 7.3h recorded data of up to 16 persons at the same time in four kitchen scenes, with more than 4M annotated human poses, represented by a parametric 3D body model. In addition, dynamic scene geometry and objects like chair or cupboard are annotated per frame. As first benchmarks, we propose two protocols for short-term and long-term human motion forecasting.

One of the most relevant tasks for augmented and virtual reality applications is the interaction of virtual objects with real humans which requires accurate 3D human pose predictions. Obtaining accurate 3D human poses requires careful camera calibration which is difficult for nontechnical personal or in a pop-up scenario. Recent markerless motion capture approaches require accurate camera calibration at least for the final triangulation step. Instead, we solve this problem by presenting ElliPose, Stereoscopic 3D Human Pose Estimation by Fitting Ellipsoids, where we jointly estimate the 3D human as well as the camera pose. We exploit the fact that bones do not change in length over the course of a sequence and thus their relative trajectories have to lie on the surface of a sphere which we can utilize to iteratively correct the camera and 3D pose estimation. As another use-case we demonstrate that our approach can be used as replacement for ground-truth 3D poses to train monocular 3D pose estimators. We show that our method produces competitive results even when comparing with state-of-the-art methods that use more cameras or ground-truth camera extrinsics.

Joint synthesis of images and segmentation masks with generative adversarial networks (GANs) is promising to reduce the effort needed for collecting image data with pixel-wise annotations. However, to learn high-fidelity image-mask synthesis, existing GAN approaches first need a pre-training phase requiring large amounts of image data, which limits their utilization in restricted image domains. In this work, we take a step to reduce this limitation, introducing the task of one-shot image-mask synthesis. We aim to generate diverse images and their segmentation masks given only a single labelled example, and assuming, contrary to previous models, no access to any pre-training data. To this end, inspired by the recent architectural developments of single-image GANs, we introduce our OSMIS model which enables the synthesis of segmentation masks that are precisely aligned to the generated images in the one-shot regime. Besides achieving the high fidelity of generated masks, OSMIS outperforms state-of-the-art single-image GAN models in image synthesis quality and diversity. In addition, despite not using any additional data, OSMIS demonstrates an impressive ability to serve as a source of useful data augmentation for one-shot segmentation applications, providing performance gains that are complementary to standard data augmentation techniques. Code is available at https://github.com/boschresearch/one-shot-synthesis.

in "A Realism Metric for Generated LiDAR Point Clouds" by Triess et al. In this work, a new metric is proposed that measures the quality of LiDAR point clouds that are generated, e.g., by a generative network. The metric can be used as an early indicator which assesses whether the generated training data will improve a down-stream task like semantic point cloud segmentation. Brissman et al. propose in "Recurrent Graph Neural Networks for Video Instance Segmentation" a very efficient approach for video instance segmentation. The approach tracks and segments multiple objects on-line and in real-time. Temporal information is also used in "Spatio-Temporal Outdoor Lighting Aggregation on Image Sequences using Transformer Networks" by Lee et al. to estimate outdoor illumination more consistently and without the need of an additional post-processing step.

In remote sensing, plenty of multispectral images are publicly available from various landcover satellite missions. Contrastive self-supervised learning is commonly applied to unlabeled data but relies on domain-specific transformations used for learning. When focusing on vegetation, standard transformations from image processing cannot be applied to the near-infrared (NIR) channel, which carries valuable information about the vegetation state. Therefore, we use contrastive learning, relying on different views of unlabeled, multispectral images to obtain a pretrained model to improve the accuracy scores on small-sized remote sensing datasets. This study presents the generation of additional views tailored to remote sensing images using atmospheric correction as an alternative transformation to color jittering. The purpose of the atmospheric transformation is to provide a physically consistent transformation. The proposed transformation can be easily integrated with multiple channels to exploit spectral signatures of objects. Our approach can be applied to other remote sensing tasks. Using this transformation leads to improved classification accuracy of up to 6%.

With the success of deep learning in classifying short trimmed videos, more attention has been focused on temporally segmenting and classifying activities in long untrimmed videos. State-of-the-art approaches for action segmentation utilize several layers of temporal convolution and temporal pooling. Despite the capabilities of these approaches in capturing temporal dependencies, their predictions suffer from over-segmentation errors. In this paper, we propose a multi-stage architecture for the temporal action segmentation task that overcomes the limitations of the previous approaches. The first stage generates an initial prediction that is refined by the next ones. In each stage we stack several layers of dilated temporal convolutions covering a large receptive field with few parameters. While this architecture already performs well, lower layers still suffer from a small receptive field. To address this limitation, we propose a dual dilated layer that combines both large and small receptive fields. We further decouple the design of the first stage from the refining stages to address the different requirements of these stages. Extensive evaluation shows the effectiveness of the proposed model in capturing long-range dependencies and recognizing action segments. Our models achieve state-of-the-art results on three datasets: 50Salads, Georgia Tech Egocentric Activities (GTEA), and the Breakfast dataset.

Climate change is expected to intensify and increase extreme events in the weather cycle. Since this has a significant impact on various sectors of our life, recent works are concerned with identifying and predicting such extreme events from Earth observations. With respect to wildfire danger forecasting, previous deep learning approaches duplicate static variables along the time dimension and neglect the intrinsic differences between static and dynamic variables. Furthermore, most existing multi-branch architectures lose the interconnections between the branches during the feature learning stage. To address these issues, this paper proposes a 2D/3D two-branch convolutional neural network (CNN) with a Location-aware Adaptive Normalization layer (LOAN). Using LOAN as a building block, we can modulate the dynamic features conditional on their geographical locations. Thus, our approach considers feature properties as a unified yet compound 2D/3D model. Besides, we propose using the sinusoidal-based encoding of the day of the year to provide the model with explicit temporal information about the target day within the year. Our experimental results show a better performance of our approach than other baselines on the challenging FireCube dataset. The results show that location-aware adaptive feature normalization is a promising technique to learn the relation between dynamic variables and their geographic locations, which is highly relevant for areas where remote sensing data builds the basis for analysis. The source code is available at https://github.com/HakamShams/LOAN.

Video-based person re-identification~(re-ID) is an important technique in visual surveillance systems which aims to match video snippets of people captured by different cameras. Existing methods are mostly based on convolutional neural networks~(CNNs), whose building blocks either process local neighbor pixels at a time, or, when 3D convolutions are used to model temporal information, suffer from the misalignment problem caused by person movement. In this paper, we propose to overcome the limitations of normal convolutions with a human-oriented graph method. Specifically, features located at person joint keypoints are extracted and connected as a spatial-temporal graph. These keypoint features are then updated by message passing from their connected nodes with a graph convolutional network~(GCN). During training, the GCN can be attached to any CNN-based person re-ID model to assist representation learning on feature maps, whilst it can be dropped after training for better inference speed. Our method brings significant improvements over the CNN-based baseline model on the MARS dataset with generated person keypoints and a newly annotated dataset: PoseTrackReID. It also defines a new state-of-the-art method in terms of top-1 accuracy and mean average precision in comparison to prior works.

This paper introduces Ranking Info Noise Contrastive Estimation (RINCE), a new member in the family of InfoNCE losses that preserves a ranked ordering of positive samples. In contrast to the standard InfoNCE loss, which requires a strict binary separation of the training pairs into similar and dissimilar samples, RINCE can exploit information about a similarity ranking for learning a corresponding embedding space. We show that the proposed loss function learns favorable embeddings compared to the standard InfoNCE whenever at least noisy ranking information can be obtained or when the definition of positives and negatives is blurry. We demonstrate this for a supervised classification task with additional superclass labels and noisy similarity scores. Furthermore, we show that RINCE can also be applied to unsupervised training with experiments on unsupervised representation learning from videos. In particular, the embedding yields higher classification accuracy, retrieval rates and performs better on out-of-distribution detection than the standard InfoNCE loss.

Current research evaluates person search, multi-object tracking and multi-person pose estimation as separate tasks and on different datasets although these tasks are very akin to each other and comprise similar sub-tasks, e.g. person detection or appearance-based association of detected persons. Consequently, approaches on these respective tasks are eligible to complement each other. Therefore, we introduce PoseTrack21, a large-scale dataset for person search, multi-object tracking and multi-person pose tracking in real-world scenarios with a high diversity of poses. The dataset provides rich annotations like human pose annotations including annotations of joint occlusions, bounding box annotations even for small persons, and person-ids within and across video sequences. The dataset allows to evaluate multi-object tracking and multi-person pose tracking jointly with person re-identification or exploit structural knowledge of human poses to improve person search and tracking, particularly in the context of severe occlusions. With PoseTrack21, we want to encourage researchers to work on joint approaches that perform reasonably well on all three tasks.

Abstract Despite their recent successes, generative adversarial networks (GANs) for semantic image synthesis still suffer from poor image quality when trained with only adversarial supervision. Previously, additionally employing the VGG-based perceptual loss has helped to overcome this issue, significantly improving the synthesis quality, but at the same time limited the progress of GAN models for semantic image synthesis. In this work, we propose a novel, simplified GAN model, which needs only adversarial supervision to achieve high quality results. We re-design the discriminator as a semantic segmentation network, directly using the given semantic label maps as the ground truth for training. By providing stronger supervision to the discriminator as well as to the generator through spatially- and semantically-aware discriminator feedback, we are able to synthesize images of higher fidelity and with a better alignment to their input label maps, making the use of the perceptual loss superfluous. Furthermore, we enable high-quality multi-modal image synthesis through global and local sampling of a 3D noise tensor injected into the generator, which allows complete or partial image editing. We show that images synthesized by our model are more diverse and follow the color and texture distributions of real images more closely. We achieve a strong improvement in image synthesis quality over prior state-of-the-art models across the commonly used ADE20K, Cityscapes, and COCO-Stuff datasets using only adversarial supervision. In addition, we investigate semantic image synthesis under severe class imbalance and sparse annotations, which are common aspects in practical applications but were overlooked in prior works. To this end, we evaluate our model on LVIS, a dataset originally introduced for long-tailed object recognition. We thereby demonstrate high performance of our model in the sparse and unbalanced data regimes, achieved by means of the proposed 3D noise and the ability of our discriminator to balance class contributions directly in the loss function. Our code and pretrained models are available at https://github.com/boschresearch/OASIS .

Action segmentation is the task of predicting the actions for each frame of a video. As obtaining the full annotation of videos for action segmentation is expensive, weakly supervised approaches that can learn only from transcripts are appealing. In this paper, we propose a novel end-to-end approach for weakly supervised action segmentation based on a two-branch neural network. The two branches of our network predict two redundant but different representations for action segmentation and we propose a novel mutual consistency (MuCon) loss that enforces the consistency of the two redundant representations. Using the MuCon loss together with a loss for transcript prediction, our proposed approach achieves the accuracy of state-of-the-art approaches while being 14 times faster to train and 20 times faster during inference. The MuCon loss proves beneficial even in the fully supervised setting.

Recently, a few works have been proposed to model the uncertainty of the future human motion. These works do not forecast a single sequence but multiple sequences for the same observation. While these works focused on increasing the diversity, this work focuses on keeping a high quality of the forecast sequences even for very long time horizons of up to 30 seconds. In order to achieve this goal, we propose to forecast the intention of the person ahead of time. This has the advantage that the generated human motion remains goal oriented and that the motion transitions between two actions are smooth and highly realistic. We furthermore propose a new quality score for evaluation that correlates better with human perception than other metrics. The results and a user study show that our approach forecasts multiple sequences that are more plausible compared to the state-of-the-art.

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.

Temporal action segmentation approaches have been very successful recently. However, annotating videos with frame-wise labels to train such models is very expensive and time consuming. While weakly supervised methods trained using only ordered action lists require less annotation effort, the performance is still worse than fully supervised approaches. In this paper, we propose to use timestamp supervision for the temporal action segmentation task. Timestamps require a comparable annotation effort to weakly supervised approaches, and yet provide a more supervisory signal. To demonstrate the effectiveness of timestamp supervision, we propose an approach to train a segmentation model using only timestamps annotations. Our approach uses the model output and the annotated timestamps to generate frame-wise labels by detecting the action changes. We further introduce a confidence loss that forces the predicted probabilities to monotonically decrease as the distance to the timestamps increases. This ensures that all and not only the most distinctive frames of an action are learned during training. The evaluation on four datasets shows that models trained with timestamps annotations achieve comparable performance to the fully supervised approaches.

Self-supervised video representation methods typically focus on the representation of temporal attributes in videos. However, the role of stationary versus non-stationary attributes is less explored: Stationary features, which remain similar throughout the video, enable the prediction of video-level action classes. Non-stationary features, which represent temporally varying attributes, are more beneficial for downstream tasks involving more fine-grained temporal understanding, such as action segmentation. We argue that a single representation to capture both types of features is sub-optimal, and propose to decompose the representation space into stationary and non-stationary features via contrastive learning from long and short views, i.e. long video sequences and their shorter sub-sequences. Stationary features are shared between the short and long views, while non-stationary features aggregate the short views to match the corresponding long view. To empirically verify our approach, we demonstrate that our stationary features work particularly well on an action recognition downstream task, while our non-stationary features perform better on action segmentation. Furthermore, we analyse the learned representations and find that stationary features capture more temporally stable, static attributes, while non-stationary features encompass more temporally varying ones.

Trajectory forecasting is a crucial step for autonomous vehicles and mobile robots in order to navigate and interact safely. In order to handle the spatial interactions between objects, graph-based approaches have been proposed. These methods, however, model motion on a frame-to-frame basis and do not provide a strong temporal model. To overcome this limitation, we propose a compact model called Spatial-Temporal Consistency Network (STC-Net). In STC-Net, dilated temporal convolutions are introduced to model long-range dependencies along each trajectory for better temporal modeling while graph convolutions are employed to model the spatial interaction among different trajectories. Furthermore, we propose a feature-wise convolution to generate the predicted trajectories in one pass and refine the forecast trajectories together with the reconstructed observed trajectories. We demonstrate that STC-Net generates spatially and temporally consistent trajectories and outperforms other graph-based methods. Since STC-Net requires only 0.7k parameters and forecasts the future with a latency of only 1.3ms, it advances the state-of-the-art and satisfies the requirements for realistic applications.

Execution monitoring is essential for robots to detect and respond to failures. Since it is impossible to enumerate all failures for a given task, we learn from successful executions of the task to detect visual anomalies during runtime. Our method learns to predict the motions that occur during the nominal execution of a task, including camera and robot body motion. A probabilistic U-Net architecture is used to learn to predict optical flow, and the robot’s kinematics and 3D model are used to model camera and body motion. The errors between the observed and predicted motion are used to calculate an anomaly score. We evaluate our method on a dataset of a robot placing a book on a shelf, which includes anomalies such as falling books, camera occlusions, and robot disturbances. We find that modeling camera and body motion, in addition to the learning-based optical flow prediction, results in an improvement of the area under the receiver operating characteristic curve from 0.752 to 0.804, and the area under the precision-recall curve from 0.467 to 0.549.

This paper introduces a novel method for self-supervised video representation learning via feature prediction. In contrast to the previous methods that focus on future feature prediction, we argue that a supervisory signal arising from unobserved past frames is complementary to one that originates from the future frames. The rationale behind our method is to encourage the network to explore the temporal structure of videos by distinguishing between future and past given present observations. We train our model in a contrastive learning framework, where joint encoding of future and past provides us with a comprehensive set of temporal hard negatives via swapping. We empirically show that utilizing both signals enriches the learned representations for the downstream task of action recognition. It out-performs independent prediction of future and past.

Codec Avatars are a recent class of learned, photorealistic face models that accurately represent the geometry and texture of a person in 3D (i.e., for virtual reality), and are almost indistinguishable from video [28]. In this paper we describe the first approach to animate these parametric models in real-time which could be deployed on commodity virtual reality hardware using audio and/or eye tracking. Our goal is to display expressive conversations between individuals that exhibit important social signals such as laughter and excitement solely from la-tent cues in our lossy input signals. To this end we collected over 5 hours of high frame rate 3D face scans across three participants including traditional neutral speech as well as expressive and conversational speech. We investigate a multimodal fusion approach that dynamically identifies which sensor encoding should animate which parts of the face at any time. See the supplemental video which demonstrates our ability to generate full face motion far beyond the typically neutral lip articulations seen in competing work: https://research.fb.com/videos/audio-and-gaze-driven-facial-animation-of-codec-avatars/.

Temporal action segmentation is a topic of increasing interest, however, annotating each frame in a video is cumbersome and costly. Weakly supervised approaches therefore aim at learning temporal action segmentation from videos that are only weakly labeled. In this work, we assume that for each training video only the list of actions is given that occur in the video, but not when, how often, and in which order they occur. In order to address this task, we propose an approach that can be trained end-to-end on such data. The approach divides the video into smaller temporal regions and predicts for each region the action label and its length. In addition, the network estimates the action labels for each frame. By measuring how consistent the frame-wise predictions are with respect to the temporal regions and the annotated action labels, the network learns to divide a video into class-consistent regions. We evaluate our approach on three datasets where the approach achieves state-of-the-art results.

Since annotating and curating large datasets is very expensive, there is a need to transfer the knowledge from existing annotated datasets to unlabelled data. Data that is relevant for a specific application, however, usually differs from publicly available datasets since it is sampled from a different domain. While domain adaptation methods compensate for such a domain shift, they assume that all categories in the target domain are known and match the categories in the source domain. Since this assumption is violated under real-world conditions, we propose an approach for open set domain adaptation where the target domain contains instances of categories that are not present in the source domain. The proposed approach achieves state-of-the-art results on various datasets for image classification and action recognition. Since the approach can be used for open set and closed set domain adaptation, as well as unsupervised and semi-supervised domain adaptation, it is a versatile tool for many applications.

Action recognition has become a rapidly developing research field within the last decade. But with the increasing demand for large scale data, the need of hand annotated data for the training becomes more and more impractical. One way to avoid frame-based human annotation is the use of action order information to learn the respective action classes. In this context, we propose a hierarchical approach to address the problem of weakly supervised learning of human actions from ordered action labels by structuring recognition in a coarse-to-fine manner. Given a set of videos and an ordered list of the occurring actions, the task is to infer start and end frames of the related action classes within the video and to train the respective action classifiers without any need for hand labeled frame boundaries. We address this problem by combining a framewise RNN model with a coarse probabilistic inference. This combination allows for the temporal alignment of long sequences and thus, for an iterative training of both elements. While this system alone already generates good results, we show that the performance can be further improved by approximating the number of subactions to the characteristics of the different action classes as well as by the introduction of a regularizing length prior. The proposed system is evaluated on two benchmark datasets, the Breakfast and the Hollywood extended dataset, showing a competitive performance on various weak learning tasks such as temporal action segmentation and action alignment.

Temporally locating and classifying action segments in long untrimmed videos is of particular interest to many applications like surveillance and robotics. While traditional approaches follow a two-step pipeline, by generating frame-wise probabilities and then feeding them to high-level temporal models, recent approaches use temporal convolutions to directly classify the video frames. In this paper, we introduce a multi-stage architecture for the temporal action segmentation task. Each stage features a set of dilated temporal convolutions to generate an initial prediction that is refined by the next one. This architecture is trained using a combination of a classification loss and a proposed smoothing loss that penalizes over-segmentation errors. Extensive evaluation shows the effectiveness of the proposed model in capturing long-range dependencies and recognizing action segments. Our model achieves state-of-the-art results on three challenging datasets: 50Salads, Georgia Tech Egocentric Activities (GTEA), and the Breakfast dataset.

The task of temporally detecting and segmenting actions in untrimmed videos has seen an increased attention recently. One problem in this context arises from the need to define and label action boundaries to create annotations for training which is very time and cost intensive. To address this issue, we propose an unsupervised approach for learning action classes from untrimmed video sequences. To this end, we use a continuous temporal embedding of framewise features to benefit from the sequential nature of activities. Based on the latent space created by the embedding, we identify clusters of temporal segments across all videos that correspond to semantic meaningful action classes. The approach is evaluated on three challenging datasets, namely the Breakfast dataset, YouTube Instructions, and the 50Salads dataset. While previous works assumed that the videos contain the same high level activity, we furthermore show that the proposed approach can also be applied to a more general setting where the content of the videos is unknown.

When humans have to solve everyday tasks, they simply pick the objects that are most suitable. While the question which object should one use for a specific task sounds trivial for humans, it is very difficult to answer for robots or other autonomous systems. This issue, however, is not addressed by current benchmarks for object detection that focus on detecting object categories. We therefore introduce the COCO-Tasks dataset which comprises about 40,000 images where the most suitable objects for 14 tasks have been annotated. We furthermore propose an approach that detects the most suitable objects for a given task. The approach builds on a Gated Graph Neural Network to exploit the appearance of each object as well as the global context of all present objects in the scene. In our experiments, we show that the proposed approach outperforms other approaches that are evaluated on the dataset like classification or ranking approaches.

Semantic scene understanding is important for various applications. In particular, self-driving cars need a fine-grained understanding of the surfaces and objects in their vicinity. Light detection and ranging (LiDAR) provides precise geometric information about the environment and is thus a part of the sensor suites of almost all self-driving cars. Despite the relevance of semantic scene understanding for this application, there is a lack of a large dataset for this task which is based on an automotive LiDAR. In this paper, we introduce a large dataset to propel research on laser-based semantic segmentation. We annotated all sequences of the KITTI Vision Odometry Benchmark and provide dense point-wise annotations for the complete 360-degree field-of-view of the employed automotive LiDAR. We propose three benchmark tasks based on this dataset: (i) semantic segmentation of point clouds using a single scan, (ii) semantic segmentation using multiple past scans, and (iii) semantic scene completion, which requires to anticipate the semantic scene in the future. We provide baseline experiments and show that there is a need for more sophisticated models to efficiently tackle these tasks. Our dataset opens the door for the development of more advanced methods, but also provides plentiful data to investigate new research directions.

We propose a Generative Adversarial Network (GAN) to forecast 3D human motion given a sequence of past 3D skeleton poses. While recent GANs have shown promising results, they can only forecast plausible motion over relatively short periods of time (few hundred milliseconds) and typically ignore the absolute position of the skeleton w.r.t. the camera. Our scheme provides long term predictions (two seconds or more) for both the body pose and its absolute position. Our approach builds upon three main contributions. First, we represent the data using a spatio-temporal tensor of 3D skeleton coordinates which allows formulating the prediction problem as an inpainting one, for which GANs work particularly well. Secondly, we design an architecture to learn the joint distribution of body poses and global motion, capable to hypothesize large chunks of the input 3D tensor with missing data. And finally, we argue that the L2 metric, considered so far by most approaches, fails to capture the actual distribution of long-term human motion. We propose two alternative metrics, based on the distribution of frequencies, that are able to capture more realistic motion patterns. Extensive experiments demonstrate our approach to significantly improve the state of the art, while also handling situations in which past observations are corrupted by occlusions, noise and missing frames.

In this paper, we propose an approach that spatially lo-calizes the activities in a video frame where each person can perform multiple activities at the same time. Our approach takes the temporal scene context as well as the relations of the actions of detected persons into account. While the temporal context is modeled by a temporal recurrent neural network (RNN), the relations of the actions are modeled by a graph RNN. Both networks are trained together and the proposed approach achieves state of the art results on the AVA dataset.

We address the task of 3D semantic scene completion, i.e., given a single depth image, we predict the semantic labels and occupancy of voxels in a 3D grid representing the scene. In light of the recently introduced generative adversarial networks (GAN), our goal is to explore the potential of this model and the efficiency of various important design choices. Our results show that using conditional GANs outperforms the vanilla GAN setup. We evaluate these architecture designs on several datasets. Based on our experiments, we demonstrate that GANs are able to outperform the performance of a baseline 3D CNN in case of clean annotations, but they suffer from poorly aligned annotations.

In this work, we address the problem how a network for action recognition that has been trained on a modality like RGB videos can be adapted to recognize actions for another modality like sequences of 3D human poses. To this end, we extract the knowledge of the trained teacher network for the source modality and transfer it to a small ensemble of student networks for the target modality. For the cross-modal knowledge distillation, we do not require any annotated data. Instead we use pairs of sequences of both modalities as supervision, which are straightforward to acquire. In contrast to previous works for knowledge distillation that use a KL-loss, we show that the cross-entropy loss together with mutual learning of a small ensemble of student networks performs better. In fact, the proposed approach for cross-modal knowledge distillation nearly achieves the accuracy of a student network trained with full supervision.

Existing systems for video-based pose estimation and tracking struggle to perform well on realistic videos with multiple people and often fail to output body-pose trajectories consistent over time. To address this shortcoming this paper introduces PoseTrack which is a new large-scale benchmark for video-based human pose estimation and articulated tracking. Our new benchmark encompasses three tasks focusing on i) single-frame multi-person pose estimation, ii) multi-person pose estimation in videos, and iii) multi-person articulated tracking. To establish the benchmark, we collect, annotate and release a new dataset that features videos with multiple people labeled with person tracks and articulated pose. A public centralized evaluation server is provided to allow the research community to evaluate on a held-out test set. Furthermore, we conduct an extensive experimental study on recent approaches to articulated pose tracking and provide analysis of the strengths and weaknesses of the state of the art. We envision that the proposed benchmark will stimulate productive research both by providing a large and representative training dataset as well as providing a platform to objectively evaluate and compare the proposed methods. The benchmark is freely accessible at https://posetrack.net/.

Analyzing human actions in videos has gained increased attention recently. While most works focus on classifying and labeling observed video frames or anticipating the very recent future, making long-term predictions over more than just a few seconds is a task with many practical applications that has not yet been addressed. In this paper, we propose two methods to predict a considerably large amount of future actions and their durations. Both, a CNN and an RNN are trained to learn future video labels based on previously seen content. We show that our methods generate accurate predictions of the future even for long videos with a huge amount of different actions and can even deal with noisy or erroneous input information.

Action detection and temporal segmentation of actions in videos are topics of increasing interest. While fully supervised systems have gained much attention lately, full annotation of each action within the video is costly and impractical for large amounts of video data. Thus, weakly supervised action detection and temporal segmentation methods are of great importance. While most works in this area assume an ordered sequence of occurring actions to be given, our approach only uses a set of actions. Such action sets provide much less supervision since neither action ordering nor the number of action occurrences are known. In exchange, they can be easily obtained, for instance, from meta-tags, while ordered sequences still require human annotation. We introduce a system that automatically learns to temporally segment and label actions in a video, where the only supervision that is used are action sets. An evaluation on three datasets shows that our method still achieves good results although the amount of supervision is significantly smaller than for other related methods.

Video learning is an important task in computer vision and has experienced increasing interest over the recent years. Since even a small amount of videos easily comprises several million frames, methods that do not rely on a frame-level annotation are of special importance. In this work, we propose a novel learning algorithm with a Viterbi-based loss that allows for online and incremental learning of weakly annotated video data. We moreover show that explicit context and length modeling leads to huge improvements in video segmentation and labeling tasks and include these models into our framework. On several action segmentation benchmarks, we obtain an improvement of up to 10% compared to current state-of-the-art methods.

A group of persons can be analyzed at various semantic levels such as individual actions, their interactions, and the activity of the entire group. In this paper, we propose a structural recurrent neural network (SRNN) that uses a series of interconnected RNNs to jointly capture the actions of individuals, their interactions, as well as the group activity. While previous structural recurrent neural networks assumed that the number of nodes and edges is constant, we use a grid pooling layer to address the fact that the number of individuals in a group can vary. We evaluate two variants of the structural recurrent neural network on the Volleyball Dataset.

In this work, we introduce the challenging problem of joint multi-person pose estimation and tracking of an unknown number of persons in unconstrained videos. Existing methods for multi-person pose estimation in images cannot be applied directly to this problem, since it also requires to solve the problem of person association over time in addition to the pose estimation for each person. We therefore propose a novel method that jointly models multi-person pose estimation and tracking in a single formulation. To this end, we represent body joint detections in a video by a spatio-temporal graph and solve an integer linear program to partition the graph into sub-graphs that correspond to plausible body pose trajectories for each person. The proposed approach implicitly handles occlusion and truncation of persons. Since the problem has not been addressed quantitatively in the literature, we introduce a challenging Multi-Person PoseTrack dataset, and also propose a completely unconstrained evaluation protocol that does not make any assumptions about the scale, size, location or the number of persons. Finally, we evaluate the proposed approach and several baseline methods on our new dataset.

We present an approach for weakly supervised learning of human actions. Given a set of videos and an ordered list of the occurring actions, the goal is to infer start and end frames of the related action classes within the video and to train the respective action classifiers without any need for hand labeled frame boundaries. To address this task, we propose a combination of a discriminative representation of subactions, modeled by a recurrent neural network, and a coarse probabilistic model to allow for a temporal alignment and inference over long sequences. While this system alone already generates good results, we show that the performance can be further improved by approximating the number of subactions to the characteristics of the different action classes. To this end, we adapt the number of subaction classes by iterating realignment and reestimation during training. The proposed system is evaluated on two benchmark datasets, the Breakfast and the Hollywood extended dataset, showing a competitive performance on various weak learning tasks such as temporal action segmentation and action alignment.

Localizing functional regions of objects or affordances is an important aspect of scene understanding and relevant for many robotics applications. In this work, we introduce a pixel-wise annotated affordance dataset of 3090 images containing 9916 object instances. Since parts of an object can have multiple affordances, we address this by a convolutional neural network for multilabel affordance segmentation. We also propose an approach to train the network from very few keypoint annotations. Our approach achieves a higher affordance detection accuracy than other weakly supervised methods that also rely on keypoint annotations or image annotations as weak supervision.

In this work we propose to utilize information about human actions to improve pose estimation in monocular videos. To this end, we present a pictorial structure model that exploits high-level information about activities to incorporate higher-order part dependencies by modeling action specific appearance models and pose priors. However, instead of using an additional expensive action recognition framework, the action priors are efficiently estimated by our pose estimation framework. This is achieved by starting with a uniform action prior and updating the action prior during pose estimation. We also show that learning the right amount of appearance sharing among action classes improves the pose estimation. We demonstrate the effectiveness of the proposed method on two challenging datasets for pose estimation and action recognition with over 80,000 test images.

When the training and the test data belong to different domains, the accuracy of an object classifier is significantly reduced. Therefore, several algorithms have been proposed in the last years to diminish the so called domain shift between datasets. However, all available evaluation protocols for domain adaptation describe a closed set recognition task, where both domains, namely source and target, contain exactly the same object classes. In this work, we also explore the field of domain adaptation in open sets, which is a more realistic scenario where only a few categories of interest are shared between source and target data. Therefore, we propose a method that fits in both closed and open set scenarios. The approach learns a mapping from the source to the target domain by jointly solving an assignment problem that labels those target instances that potentially belong to the categories of interest present in the source dataset. A thorough evaluation shows that our approach outperforms the state-of-the-art.

This paper proposes an end-to-end learning framework for multiview stereopsis. We term the network SurfaceNet. It takes a set of images and their corresponding camera parameters as input and directly infers the 3D model. The key advantage of the framework is that both photo-consistency as well geometric relations of the surface structure can be directly learned for the purpose of multiview stereopsis in an end-to-end fashion. SurfaceNet is a fully 3D convolutional network which is achieved by encoding the camera parameters together with the images in a 3D voxel representation. We evaluate SurfaceNet on the large-scale DTU benchmark.

While current approaches to action recognition on presegmented video clips already achieve high accuracies, temporal action detection is still far from comparably good results. Automatically locating and classifying the relevant action segments in videos of varying lengths proves to be a challenging task. We propose a novel method for temporal action detection including statistical length and language modeling to represent temporal and contextual structure. Our approach aims at globally optimizing the joint probability of three components, a length and language model and a discriminative action model, without making intermediate decisions. The problem of finding the most likely action sequence and the corresponding segment boundaries in an exponentially large search space is addressed by dynamic programming. We provide an extensive evaluation of each model component on Thumos 14, a large action detection dataset, and report state-of-the-art results on three datasets.

One major challenge for 3D pose estimation from a single RGB image is the acquisition of sufficient training data. In particular, collecting large amounts of training data that contain unconstrained images and are annotated with accurate 3D poses is infeasible. We therefore propose to use two independent training sources. The first source consists of images with annotated 2D poses and the second source consists of accurate 3D motion capture data. To integrate both sources, we propose a dual-source approach that combines 2D pose estimation with efficient and robust 3D pose retrieval. In our experiments, we show that our approach achieves state-of-the-art results and is even competitive when the skeleton structure of the two sources differ substantially.

We describe an end-to-end generative approach for the segmentation and recognition of human activities. In this approach, a visual representation based on reduced Fisher Vectors is combined with a structured temporal model for recognition. We show that the statistical properties of Fisher Vectors make them an especially suitable front-end for generative models such as Gaussian mixtures. The system is evaluated for both the recognition of complex activities as well as their parsing into action units. Using a variety of video datasets ranging from human cooking activities to animal behaviors, our experiments demonstrate that the resulting architecture outperforms state-of-the-art approaches for larger datasets, i.e. when sufficient amount of data is available for training structured generative models.

Large image datasets such as ImageNet or open-ended photo websites like Flickr are revealing new challenges to image classification that were not apparent in smaller, fixed sets. In particular, the efficient handling of dynamically growing datasets, where not only the amount of training data but also the number of classes increases over time, is a relatively unexplored problem. In this challenging setting, we study how two variants of Random Forests (RF) perform under four strategies to incorporate new classes while avoiding to retrain the RFs from scratch. The various strategies account for different trade-offs between classification accuracy and computational efficiency. In our extensive experiments, we show that both RF variants, one based on Nearest Class Mean classifiers and the other on SVMs, outperform conventional RFs and are well suited for incrementally learning new classes. In particular, we show that RFs initially trained with just 10 classes can be extended to 1,000 classes with an acceptable loss of accuracy compared to training from the full data and with great computational savings compared to retraining for each new batch of classes.

Bag-of-words (BoW) models are widely used in the field of computer vision. A BoW model consists of a visual vocabulary that is generated by unsupervised clustering the features of the training data, e.g., by using kMeans. The clustering methods, however, struggle with large amounts of data, in particular, in the context of action recognition. In this paper, we propose a transformation of the standard BoW model into a neural network, enabling discriminative training of the visual vocabulary on large action recog-nition datasets. We show that our model is equivalent to the original BoW model but allows for the application of supervised neural network training. Our model outperforms the conventional BoW model and sparse coding methods on recent action recognition benchmarks. 1

The number of digital images is growing extremely rapidly, and so is the need for their classification. But, as more images of pre-defined categories become available, they also become more diverse and cover finer semantic differences. Ultimately, the categories themselves need to be divided into subcategories to account for that semantic refinement. Image classification in general has improved significantly over the last few years, but it still requires a massive amount of manually annotated data. Subdividing categories into subcategories multiples the number of labels, aggravating the annotation problem. Hence, we can expect the annotations to be refined only for a subset of the already labeled data, and exploit coarser labeled data to improve classification. In this work, we investigate how coarse category labels can be used to improve the classification of subcategories. To this end, we adopt the framework of Random Forests and propose a regularized objective function that takes into account relations between categories and subcategories. Compared to approaches that disregard the extra coarse labeled data, we achieve a relative improvement in subcategory classification accuracy of up to 22% in our large-scale image classification experiments.

Recent advances have enabled 3d object reconstruction approaches using a single off-the-shelf RGB-D camera. Although these approaches are successful for a wide range of object classes, they rely on stable and distinctive geometric or texture features. Many objects like mechanical parts, toys, household or decorative articles, however, are textureless and characterized by minimalistic shapes that are simple and symmetric. Existing in-hand scanning systems and 3d reconstruction techniques fail for such symmetric objects in the absence of highly distinctive features. In this work, we show that extracting 3d hand motion for in-hand scanning effectively facilitates the reconstruction of even featureless and highly symmetric objects and we present an approach that fuses the rich additional information of hands into a 3d reconstruction pipeline, significantly contributing to the state-of-the-art of in-hand scanning.

In recent years, large image data sets such as "ImageNet", "TinyImages" or ever-growing social networks like "Flickr" have emerged, posing new challenges to image classification that were not apparent in smaller image sets. In particular, the efficient handling of dynamically growing data sets, where not only the amount of training images, but also the number of classes increases over time, is a relatively unexplored problem. To remedy this, we introduce Nearest Class Mean Forests (NCMF), a variant of Random Forests where the decision nodes are based on nearest class mean (NCM) classification. NCMFs not only outperform conventional random forests, but are also well suited for integrating new classes. To this end, we propose and compare several approaches to incorporate data from new classes, so as to seamlessly extend the previously trained forest instead of re-training them from scratch. In our experiments, we show that NCMFs trained on small data sets with 10 classes can be extended to large data sets with 1000 classes without significant loss of accuracy compared to training from scratch on the full data.

Hough-based voting approaches have been successfully applied to object detection. While these methods can be efficiently implemented by random forests, they estimate the probability for an object hypothesis independently for each feature. In this work, we address this problem by grouping features in a local neighborhood to obtain a better estimate of the probability. To this end, we propose oblique classification-regression forests that combine features of different trees. We further investigate the benefit of combining independent and grouped features and evaluate the approach on RGB and RGB-D datasets.

In this work, we address the problem of estimating 2d human pose from still images. Articulated body pose estimation is challenging due to the large variation in body poses and appearances of the different body parts. Recent methods that rely on the pictorial structure framework have shown to be very successful in solving this task. They model the body part appearances using discriminatively trained, independent part templates and the spatial relations of the body parts using a tree model. Within such a framework, we address the problem of obtaining better part templates which are able to handle a very high variation in appearance. To this end, we introduce parts dependent body joint regressors which are random forests that operate over two layers. While the first layer acts as an independent body part classifier, the second layer takes the estimated class distributions of the first one into account and is thereby able to predict joint locations by modeling the interdependence and co-occurrence of the parts. This helps to overcome typical ambiguities of tree structures, such as self-similarities of legs and arms. In addition, we introduce a novel data set termed FashionPose that contains over 7,000 images with a challenging variation of body part appearances due to a large variation of dressing styles. In the experiments, we demonstrate that the proposed parts dependent joint regressors outperform independent classifiers or regressors. The method also performs better or similar to the state-of-the-art in terms of accuracy, while running with a couple of frames per second.

In this work, we address the problem of estimating 2d human pose from still images. Recent methods that rely on discriminatively trained deformable parts organized in a tree model have shown to be very successful in solving this task. Within such a pictorial structure framework, we address the problem of obtaining good part templates by proposing novel, non-linear joint regressors. In particular, we employ two-layered random forests as joint regressors. The first layer acts as a discriminative, independent body part classifier. The second layer takes the estimated class distributions of the first one into account and is thereby able to predict joint locations by modeling the interdependence and co-occurrence of the parts. This results in a pose estimation framework that takes dependencies between body parts already for joint localization into account and is thus able to circumvent typical ambiguities of tree structures, such as for legs and arms. In the experiments, we demonstrate that our body parts dependent joint regressors achieve a higher joint localization accuracy than tree-based state-of-the-art methods.

Although action recognition in videos is widely studied, current methods often fail on real-world datasets. Many recent approaches improve accuracy and robustness to cope with challenging video sequences, but it is often unclear what affects the results most. This paper attempts to provide insights based on a systematic performance evaluation using thoroughly-annotated data of human actions. We annotate human Joints for the HMDB dataset (J-HMDB). This annotation can be used to derive ground truth optical flow and segmentation. We evaluate current methods using this dataset and systematically replace the output of various algorithms with ground truth. This enables us to discover what is important - for example, should we work on improving flow algorithms, estimating human bounding boxes, or enabling pose estimation? In summary, we find that high-level pose features greatly outperform low/mid level features, in particular, pose over time is critical, but current pose estimation algorithms are not yet reliable enough to provide this information. We also find that the accuracy of a top-performing action recognition framework can be greatly increased by refining the underlying low/mid level features, this suggests it is important to improve optical flow and human detection algorithms. Our analysis and J-HMDB dataset should facilitate a deeper understanding of action recognition algorithms.

Capturing the skeleton motion and detailed time-varying surface geometry of multiple, closely interacting peoples is a very challenging task, even in a multicamera setup, due to frequent occlusions and ambiguities in feature-to-person assignments. To address this task, we propose a framework that exploits multiview image segmentation. To this end, a probabilistic shape and appearance model is employed to segment the input images and to assign each pixel uniquely to one person. Given the articulated template models of each person and the labeled pixels, a combined optimization scheme, which splits the skeleton pose optimization problem into a local one and a lower dimensional global one, is applied one by one to each individual, followed with surface estimation to capture detailed nonrigid deformations. We show on various sequences that our approach can capture the 3D motion of humans accurately even if they move rapidly, if they wear wide apparel, and if they are engaged in challenging multiperson motions, including dancing, wrestling, and hugging.

Temporal clustering of human motion into semantically meaningful behaviors is a challenging task. While unsupervised methods do well to some extent, the obtained clusters often lack a semantic interpretation. In this paper, we propose to learn what makes a sequence of human poses different from others such that it should be annotated as an action. To this end, we formulate the problem as weakly supervised temporal clustering for an unknown number of clusters. Weak supervision is attained by learning a metric from the implicit semantic distances derived from already annotated databases. Such a metric contains some low-level semantic information that can be used to effectively segment a human motion sequence into distinct actions or behaviors. The main advantage of our approach is that metrics can be successfully used across datasets, making our method a compelling alternative to unsupervised methods. Experiments on publicly available mocap datasets show the effectiveness of our approach.

Hough transform based object detectors divide an object into a number of patches and combine them using a shape model. For efficient combination of patches into the shape model, the individual patches are assumed to be independent of one another. Although this independence assumption is key for fast inference, it requires the individual patches to have a high discriminative power in predicting the class and location of objects. In this paper, we argue that the sparsity of the appearance of a patch in its neighborhood can be a very powerful measure to increase the discriminative power of a local patch and incorporate it as a sparsity potential for object detection. Further, we show that this potential shall depend on the appearance of the patch to adapt to the statistics of the neighborhood specific to the type of appearance (e.g. texture or structure) it represents. We have evaluated our method on challenging datasets including the PASCAL VOC 2007 dataset and show that using the proposed sparsity potential result in a substantial improvement in the detection accuracy.

Although facial feature detection from 2D images is a well-studied field, there is a lack of real-time methods that estimate feature points even on low quality images. Here we propose conditional regression forest for this task. While regression forest learn the relations between facial image patches and the location of feature points from the entire set of faces, conditional regression forest learn the relations conditional to global face properties. In our experiments, we use the head pose as a global property and demonstrate that conditional regression forests outperform regression forests for facial feature detection. We have evaluated the method on the challenging Labeled Faces in the Wild [20] database where close-to-human accuracy is achieved while processing images in real-time.

In recent years, the rise of digital image and video data available has led to an increasing demand for image annotation. In this paper, we propose an interactive object annotation method that incrementally trains an object detector while the user provides annotations. In the design of the system, we have focused on minimizing human annotation time rather than pure algorithm learning performance. To this end, we optimize the detector based on a realistic annotation cost model based on a user study. Since our system gives live feedback to the user by detecting objects on the fly and predicts the potential annotation costs of unseen images, data can be efficiently annotated by a single user without excessive waiting time. In contrast to popular tracking-based methods for video annotation, our method is suitable for both still images and video. We have evaluated our interactive annotation approach on three datasets, ranging from surveillance, television, to cell microscopy.

Early works on human action recognition focused on tracking and classifying articulated body motions. Such methods required accurate localisation of body parts, which is a difficult task, particularly under realistic imaging conditions. As such, recent trends have shifted towards the use of more abstract, low-level appearance features such as spatio-temporal interest points. Motivated by the recent progress in pose estimation, we feel that pose-based action recognition systems warrant a second look. In this paper, we address the question of whether pose estimation is useful for action recognition or if it is better to train a classifier only on low-level appearance features drawn from video data. We compare pose-based, appearance-based and combined pose and appearance features for action recognition in a home-monitoring scenario. Our experiments show that pose-based features outperform low-level appearance features, even when heavily corrupted by noise, suggesting that pose estimation is beneficial for the action recognition task. © 2011. The copyright of this document resides with its authors.

Fast and reliable algorithms for estimating the head pose are essential for many applications and higher-level face analysis tasks. We address the problem of head pose estimation from depth data, which can be captured using the ever more affordable 3D sensing technologies available today. To achieve robustness, we formulate pose estimation as a regression problem. While detecting specific face parts like the nose is sensitive to occlusions, learning the regression on rather generic surface patches requires enormous amount of training data in order to achieve accurate estimates. We propose to use random regression forests for the task at hand, given their capability to handle large training datasets. Moreover, we synthesize a great amount of annotated training data using a statistical model of the human face. In our experiments, we show that our approach can handle real data presenting large pose changes, partial occlusions, and facial expressions, even though it is trained only on synthetic neutral face data. We have thoroughly evaluated our system on a publicly available database on which we achieve state-of-the-art performance without having to resort to the graphics card.

Unsupervised categorization of objects is a fundamental problem in computer vision. While appearance-based methods have become popular recently, other important cues like functionality are largely neglected. Motivated by psycho logical studies giving evidence that human demonstration has a facilitative effect on categorization in infancy, we pro pose an approach for object categorization from depth video streams. To this end, we have developed a method for cap turing human motion in real-time. The captured data is then used to temporally segment the depth streams into actions. The set of segmented actions are then categorized in an un supervised manner, through a novel descriptor for motion capture data that is robust to subject variations. Further more, we automatically localize the object that is manipulated within a video segment, and categorize it using the corresponding action. For evaluation, we have recorded a dataset that comprises depth data with registered video sequences for 6 subjects, 13 action classes, and 174 object manipulations.

Many object classes are primarily defined by their functions. However, this fact has been left largely unexploited by visual object categorization or detection systems. We propose a method to learn an affordance detector. It identifies locations in the 3d space which “support” the particular function. Our novel approach “imagines” an actor performing an action typical for the target object class, instead of relying purely on the visual object appearance. So, function is handled as a cue complementary to appearance, rather than being a consideration after appearance-based detection. Experimental results are given for the functional category “sitting”. Such affordance is tested on a 3d representation of the scene, as can be realistically obtained through SfM or depth cameras. In contrast to appearance-based object detectors, affordance detection requires only very few training examples and generalizes very well to other sittable objects like benches or sofas when trained on a few chairs.

We present a markerless motion capture approach that reconstructs the skeletal motion and detailed time-varying surface geometry of two closely interacting people from multi-view video. Due to ambiguities in feature-to-person assignments and frequent occlusions, it is not feasible to directly apply single-person capture approaches to the multi-person case. We therefore propose a combined image segmentation and tracking approach to overcome these difficulties. A new probabilistic shape and appearance model is employed to segment the input images and to assign each pixel uniquely to one person. Thereafter, a single-person markerless motion and surface capture approach can be applied to each individual, either one-by-one or in parallel, even under strong occlusions. We demonstrate the performance of our approach on several challenging multi-person motions, including dance and martial arts, and also provide a reference dataset for multi-person motion capture with ground truth.

Scalability of object detectors with respect to the number of classes is a very important issue for applications where many object classes need to be detected. While combining single-class detectors yields a linear complexity for testing, multi-class detectors that localize all objects at once come often at the cost of a reduced detection accuracy. In this work, we present a scalable multi-class detection algorithm which scales sublinearly with the number of classes without compromising accuracy. To this end, a shared discriminative codebook of feature appearances is jointly trained for all classes and detection is also performed for all classes jointly. Based on the learned sharing distributions of features among classes, we build a taxonomy of object classes. The taxonomy is then exploited to further reduce the cost of multi-class object detection. Our method has linear training and sublinear detection complexity in the number of classes. We have evaluated our method on the challenging PASCAL VOC'06 and PASCAL VOC'07 datasets and show that scaling the system does not lead to a loss in accuracy.

Animating hand-object interactions is a frequent task in applications such as the production of 3d movies. Unfortunately this task is difficult due to the hand's many degrees of freedom and the constraints on the hand motion imposed by the geometry of the object. However, the causality between the object state and the hand's pose can be exploited in order to simplify the animation process. In this paper, we present a method that takes an animation of an object as input and automatically generates the corresponding hand motion. This approach is based on the simple observation that objects are easier to animate than hands, since they usually have fewer degrees of freedom. The method is data-driven; sequences of hands manipulating an object are captured semi-automatically with a structured-light setup. The training data is then combined with a new animation of the object in order to generate a plausible animation featuring the hand-object interaction.

Human motion capturing (HMC) from multiview image sequences is an extremely difficult problem due to depth and orientation ambiguities and the high dimensionality of the state space. In this paper, we introduce a novel hybrid HMC system that combines video input with sparse inertial sensor input. Employing an annealing particle-based optimization scheme, our idea is to use orientation cues derived from the inertial input to sample particles from the manifold of valid poses. Then, visual cues derived from the video input are used to weight these particles and to iteratively derive the final pose. As our main contribution, we propose an efficient sampling procedure where the particles are derived analytically using inverse kinematics on the orientation cues. Additionally, we introduce a novel sensor noise model to account for uncertainties based on the von Mises-Fisher distribution. Doing so, orientation constraints are naturally fulfilled and the number of needed particles can be kept very small. More generally, our method can be used to sample poses that fulfill arbitrary orientation or positional kinematic constraints. In the experiments, we show that our system can track even highly dynamic motions in an outdoor environment with changing illumination, background clutter, and shadows.

We present an approach for modeling the human body by Sums of spatial Gaussians (SoG), allowing us to perform fast and high-quality markerless motion capture from multi-view video sequences. The SoG model is equipped with a color model to represent the shape and appearance of the human and can be reconstructed from a sparse set of images. Similar to the human body, we also represent the image domain as SoG that models color consistent image blobs. Based on the SoG models of the image and the human body, we introduce a novel continuous and differentiable model-to-image similarity measure that can be used to estimate the skeletal motion of a human at 5-15 frames per second even for many camera views. In our experiments, we show that our method, which does not rely on silhouettes or training data, offers an good balance between accuracy and computational cost.

A common approach for handling the complexity and inherent ambiguities of 3D human pose estimation is to use pose priors learned from training data. Existing approaches however, are either too simplistic (linear), too complex to learn, or can only learn latent spaces from "simple data", i.e., single activities such as walking or running. In this paper, we present an efficient stochastic gradient descent algorithm that is able to learn probabilistic non-linear latent spaces composed of multiple activities. Furthermore, we derive an incremental algorithm for the online setting which can update the latent space without extensive relearning. We demonstrate the effectiveness of our approach on the task of monocular and multi-view tracking and show that our approach outperforms the state-of-the-art.

Abstract—The paper introduces Hough forests, which are random forests adapted to perform a generalized Hough transform in an efficient way. Compared to previous Hough-based systems such as implicit shape models, Hough forests improve the performance of the generalized Hough transform for object detection on a categorical level. At the same time, their flexibility permits extensions of the Hough transform to new domains such as object tracking and action recognition. Hough forests can be regarded as task-adapted codebooks of local appearance that allow fast supervised training and fast matching at test time. They achieve high detection accuracy since the entries of such codebooks are optimized to cast Hough votes with small variance and since their efficiency permits dense sampling of local image patches or video cuboids during detection. The efficacy of Hough forests for a set of computer vision tasks is validated through experiments on a large set of publicly available benchmark data sets and comparisons with the state-of-the-art.

Off-line trained class-specific object detectors are designed to detect any instance of the class in a given image or video sequence. In the context of object tracking, however, one seeks the location and scale of a target object, which is a specific instance of the class. Hence, the target needs to be separated not only from the background but also from other instances in the video sequence. We address this problem by adapting a class-specific object detector to the target, making it more instance-specific. To this end, we learn offline a codebook for the object class that models the spatial distribution and appearance of object parts. For tracking, the codebook is coupled with a particle filter. While the posterior probability of the location and scale of the target is used to learn on-line the probability of each part in the codebook belonging to the target, the probabilistic votes for the object cast by the codebook entries are used to model the likelihood. © 2010. The copyright of this document resides with its authors.

In this paper, we propose a prior for hand pose estimation that integrates the direct relation between a manipulating hand and a 3d object. This is of particular interest for a variety of applications since many tasks performed by humans require hand-object interaction. Inspired by the ability of humans to learn the handling of an object from a single example, our focus lies on very sparse training data. We express estimated hand poses in local object coordinates and extract for each individual hand segment, the relative position and orientation as well as contact points on the object. The prior is then modeled as a spatial distribution conditioned to the object. Given a new object of the same object class and new hand dimensions, we can transfer the prior by a procedure involving a geometric warp. In our experiments, we demonstrate that the prior may be used to improve the robustness of a 3d hand tracker and to synthesize a new hand grasping a new object. For this, we integrate the prior into a unified belief propagation framework for tracking and synthesis.

We present a method to classify and localize human actions in video using a Hough transform voting framework. Random trees are trained to learn a mapping between densely-sampled feature patches and their corresponding votes in a spatio-temporal-action Hough space. The leaves of the trees form a discriminative multi-class codebook that share features between the action classes and vote for action centers in a probabilistic manner. Using low-level features such as gradients and optical flow, we demonstrate that Hough-voting can achieve state-of-the-art performance on several datasets covering a wide range of action-recognition scenarios.

Local optimization and filtering have been widely applied to model-based 3D human motion capture. Global stochastic optimization has recently been proposed as promising alternative solution for tracking and initialization. In order to benefit from optimization and filtering, we introduce a multi-layer framework that combines stochastic optimization, filtering, and local optimization. While the first layer relies on interacting simulated annealing and some weak prior information on physical constraints, the second layer refines the estimates by filtering and local optimization such that the accuracy is increased and ambiguities are resolved over time without imposing restrictions on the dynamics. In our experimental evaluation, we demonstrate the significant improvements of the multi-layer framework and provide quantitative 3D pose tracking results for the complete HumanEva-II dataset. The paper further comprises a comparison of global stochastic optimization with particle filtering, annealed particle filtering, and local optimization.

In this paper, we propose the combined use of complementary concepts for 3D tracking: region fitting on one side and dense optical flow as well as tracked SIFT features on the other. Both concepts are chosen such that they can compensate for the shortcomings of each other. While tracking by the object region can prevent the accumulation of errors, optical flow and SIFT can handle larger transformations. Whereas segmentation works best in case of homogeneous objects, optical flow computation and SIFT tracking rely on sufficiently structured objects. We show that a sensible combination yields a general tracking system that can be applied in a large variety of scenarios without the need to manually adjust weighting parameters.

Communication between humans deeply relies on the capability of expressing and recognizing feelings. For this reason, research on human-machine interaction needs to focus on the recognition and simulation of emotional states, prerequisite of which is the collection of affective corpora. Currently available datasets still represent a bottleneck for the difficulties arising during the acquisition and labeling of affective data. In this work, we present a new audio-visual corpus for possibly the two most important modalities used by humans to communicate their emotional states, namely speech and facial expression in the form of dense dynamic 3-D face geometries. We acquire high-quality data by working in a controlled environment and resort to video clips to induce affective states. The annotation of the speech signal includes: transcription of the corpus text into the phonological representation, accurate phone segmentation, fundamental frequency extraction, and signal intensity estimation of the speech signals. We employ a real-time 3-D scanner to acquire dense dynamic facial geometries and track the faces throughout the sequences, achieving full spatial and temporal correspondences. The corpus is a valuable tool for applications like affective visual speech synthesis or view-independent facial expression recognition.

We present a new performance capture approach that incorporates a physically-based cloth model to reconstruct a rigged fully-animatable virtual double of a real person in loose apparel from multi-view video recordings. Our algorithm only requires a minimum of manual interaction. Without the use of optical markers in the scene, our algorithm first reconstructs skeleton motion and detailed time-varying surface geometry of a real person from a reference video sequence. These captured reference performance data are then analyzed to automatically identify non-rigidly deforming pieces of apparel on the animated geometry. For each piece of apparel, parameters of a physically-based real-time cloth simulation model are estimated, and surface geometry of occluded body regions is approximated. The reconstructed character model comprises a skeleton-based representation for the actual body parts and a physically-based simulation model for the apparel. In contrast to previous performance capture methods, we can now also create new real-time animations of actors captured in general apparel.

We present a novel method for mouth localization in the context of multimodal speech recognition where audio and visual cues are fused to improve the speech recognition accuracy. While facial feature points like mouth corners or lip contours are commonly used to estimate at least scale, position, and orientation of the mouth, we propose a Hough transform-based method. Instead of relying on a predefined sparse subset of mouth features, it casts probabilistic votes for the mouth center from several patches in the neighborhood and accumulates the votes in a Hough image. This makes the localization more robust as it does not rely on the detection of a single feature. In addition, we exploit the different shape properties of eyes and mouth in order to localize the mouth more efficiently. Using the rotation invariant representation of the iris, scale and orientation can be efficiently inferred from the localized eye positions. The superior accuracy of our method and quantitative improvements for audio-visual speech recognition over monomodal approaches are demonstrated on two datasets. © 2009. The copyright of this document resides with its authors.

This paper proposes a method for capturing the performance of a human or an animal from a multi-view video sequence. Given an articulated template model and silhouettes from a multi-view image sequence, our approach recovers not only the movement of the skeleton, but also the possibly non-rigid temporal deformation of the 3D surface. While large scale deformations or fast movements are captured by the skeleton pose and approximate surface skinning, true small scale deformations or non-rigid garment motion are captured by fitting the surface to the silhouette. We further propose a novel optimization scheme for skeleton-based pose estimation that exploits the skeleton's tree structure to split the optimization problem into a local one and a lower dimensional global one. We show on various sequences that our approach can capture the 3D motion of animals and humans accurately even in the case of rapid movements and wide apparel like skirts.

In this work we present an approach for markerless motion capture (MoCap) of articulated objects, which are recorded with multiple unsynchronized moving cameras. Instead of using fixed (and expensive) hardware synchronized cameras, this approach allows us to track people with off-the-shelf handheld video cameras. To prepare a sequence for motion capture, we first reconstruct the static background and the position of each camera using Structure-from-Motion (SfM). Then the cameras are registered to each other using the reconstructed static background geometry. Camera synchronization is achieved via the audio streams recorded by the cameras in parallel. Finally, a markerless MoCap approach is applied to recover positions and joint configurations of subjects. Feature tracks and dense background geometry are further used to stabilize the MoCap. The experiments show examples with highly challenging indoor and outdoor scenes.

This work addresses the problem of tracking humans with skeleton-based shape models where video footage is acquired by multiple cameras. Since the shape deformations are parameterized by the skeleton, the position, orientation, and configuration of the human skeleton are estimated such that the deformed shape model is best explained by the image data. To solve this problem, several algorithms have been proposed over the last years. The approaches usually rely on filtering, local optimization, or global optimization. The global optimization algorithms can be further divided into single hypothesis (SHO) and multiple hypothesis optimization (MHO). We briefly compare the underlying mathematical models and evaluate the performance of one representative algorithm for each class. Furthermore, we compare several likelihoods and parameter settings with respect to accuracy and computation cost. A thorough evaluation is performed on two sequences with uncontrolled lighting conditions and non-static background. In addition, we demonstrate the impact of the likelihood on the HumanEva benchmark. Our results provide a guidance on algorithm design for different applications related to human motion capture.

Model-based 3D tracker estimate the position, rotation, and joint angles of a given model from video data of one or multiple cameras. They often rely on image features that are tracked over time but the accumulation of small errors results in a drift away from the target object. In this work, we address the drift problem for the challenging task of human motion capture and tracking in the presence of multiple moving objects where the error accumulation becomes even more problematic due to occlusions. To this end, we propose an analysis-by-synthesis framework for articulated models. It combines the complementary concepts of patch-based and region-based matching to track both structured and homogeneous body parts. The performance of our method is demonstrated for rigid bodies, body parts, and full human bodies where the sequences contain fast movements, self-occlusions, multiple moving objects, and clutter. We also provide a quantitative error analysis and comparison with other model-based approaches.