End-to-end autonomous driving has achieved remarkable advancements in recent years. Existing methods primarily follow a perception–planning paradigm, where perception and planning are executed sequentially within a fully differentiable framework for planning-oriented optimization. We further advance this paradigm through a "perception-in-plan'' framework design, which integrates perception into the planning process. This design facilitates targeted perception guided by evolving planning objectives over time, ultimately enhancing planning performance. Building on this insight, we introduce VeteranAD, a coupled perception and planning framework for end-to-end autonomous driving. By incorporating multi-mode anchored trajectories as planning priors, the perception module is specifically designed to gather traffic elements along these trajectories, enabling comprehensive and targeted perception. Planning trajectories are then generated based on both the perception results and the planning priors. To make perception fully serve planning, we adopt an autoregressive strategy that progressively predicts future trajectories while focusing on relevant regions for targeted perception at each step. With this simple yet effective design, VeteranAD fully unleashes the potential of planning-oriented end-to-end methods, leading to more accurate and reliable driving behavior. Extensive experiments on the NAVSIM and Bench2Drive datasets demonstrate that our VeteranAD achieves state-of-the-art performance.

The extraction of road network is essential for the generation of high-definition maps since it enables the precise localization of road landmarks and their interconnections. However, generating road network poses a significant challenge due to the conflicting underlying combination of Euclidean (e.g., road landmarks location) and non-Euclidean (e.g., road topological connectivity) structures. Existing methods struggle to merge the two types of data domains effectively, but few of them address it properly. Instead, our work establishes a unified representation of both types of data domain by projecting both Euclidean and non-Euclidean data into an integer series called RoadNet Sequence. Further than modeling an auto-regressive sequence-to-sequence Transformer model to understand RoadNet Sequence, we decouple the dependency of RoadNet Sequence into a mixture of auto-regressive and non-autoregressive dependency. Building on this, our proposed non-autoregressive sequence-to-sequence approach leverages non-autoregressive dependencies while fixing the gap towards auto-regressive dependencies, resulting in success in both efficiency and accuracy. We further identify two main bottlenecks in the current RoadNetTransformer on a non-overfitting split of the dataset: poor landmark detection limited by the BEV Encoder and error propagation to topology reasoning. Therefore, we propose Topology-Inherited Training to inherit better topology knowledge into RoadNetTransformer. Additionally, we collect SD-Maps from open-source map datasets and use this prior information to significantly improve landmark detection and reachability. Extensive experiments on the nuScenes dataset demonstrate the superiority of RoadNet Sequence representation and the non-autoregressive approach compared to existing state-of-the-art alternatives.

Point management is critical for optimizing 3D Gaussian Splatting models, as point initiation (e.g., via structure from motion) is often distributionally inappropriate. Typically, Adaptive Density Control (ADC) algorithm is adopted, leveraging view-averaged gradient magnitude thresholding for point densification, opacity thresholding for pruning, and regular all-points opacity reset. We reveal that this strategy is limited in tackling intricate/special image regions (e.g., transparent) due to inability of identifying all 3D zones requiring point densification, and lacking an appropriate mechanism to handle ill-conditioned points with negative impacts (e.g., occlusion due to false high opacity). To address these limitations, we propose a Localized Point Management (LPM) strategy, capable of identifying those error-contributing zones in greatest need for both point addition and geometry calibration. Zone identification is achieved by leveraging the underlying multiview geometry constraints, subject to image rendering errors. We apply point densification in the identified zones and then reset the opacity of the points in front of these regions, creating a new opportunity to correct poorly conditioned points. Serving as a versatile plugin, LPM can be seamlessly integrated into existing static 3D and dynamic 4D Gaussian Splatting models with minimal additional cost. Experimental evaluations validate the efficacy of our LPM in boosting a variety of existing 3D/4D models both quantitatively and qualitatively. Notably, LPM improves both static 3DGS and dynamic SpaceTimeGS to achieve state-of-the-art rendering quality while retaining real-time speeds, excelling on challenging datasets such as Tanks & Temples and the Neural 3D Video dataset.

Inverse rendering aims to recover scene geometry, material properties, and lighting from multi-view images. Given the complexity of light-surface interactions, importance sampling is essential for the evaluation of the rendering equation, as it reduces variance and enhances the efficiency of Monte Carlo sampling. Existing inverse rendering methods typically use pre-defined non-learnable importance samplers in prior manually, struggling to effectively match the spatially and directionally varied integrand and resulting in high variance and suboptimal performance. To address this limitation, we propose the concept of learning a spatially and directionally aware importance sampler for the rendering equation to accurately and flexibly capture the unconstrained complexity of a typical scene. We further formulate TensoFlow, a generic approach for sampler learning in inverse rendering, enabling to closely match the integrand of the rendering equation spatially and directionally. Concretely, our sampler is parameterized by normalizing flows, allowing both directional sampling of incident light and probability density function (PDF) inference. To capture the characteristics of the sampler spatially, we learn a tensorial representation of the scene space, which imposes spatial conditions, together with reflected direction, leading to spatially and directionally aware sampling distributions. Our model can be optimized by minimizing the difference between the integrand and our normalizing flow. Extensive experiments validate the superiority of TensoFlow over prior alternatives on both synthetic and real-world benchmarks.

In inverse rendering, accurately modeling visibility and indirect radiance for incident light is essential for capturing secondary effects. Due to the absence of a powerful Gaussian ray tracer, previous 3DGS-based methods have either adopted a simplified rendering equation or used learnable parameters to approximate incident light, resulting in inaccurate material and lighting estimations. To this end, we introduce inter-reflective Gaussian splatting (IRGS) for inverse rendering. To capture inter-reflection, we apply the full rendering equation without simplification and compute incident radiance on the fly using the proposed differentiable 2D Gaussian ray tracing. Additionally, we present an efficient optimization scheme to handle the computational demands of Monte Carlo sampling for rendering equation evaluation. Furthermore, we introduce a novel strategy for querying the indirect radiance of incident light when relighting the optimized scenes. Extensive experiments on multiple standard benchmarks validate the effectiveness of IRGS, demonstrating its capability to accurately model complex inter-reflection effects.

Generating high-fidelity, controllable, and annotated training data is critical for autonomous driving. Existing methods typically generate a single data form directly from a coarse scene layout, which not only fails to output rich data forms required for diverse downstream tasks but also struggles to model the direct layout-to-data distribution. In this paper, we introduce UniScene, the first unified framework for generating three key data forms — semantic occupancy, video, and LiDAR — in driving scenes. UniScene employs a progressive generation process that decomposes the complex task of scene generation into two hierarchical steps: (a) first generating semantic occupancy from a customized scene layout as a meta scene representation rich in both semantic and geometric information, and then (b) conditioned on occupancy, generating video and LiDAR data, respectively, with two novel transfer strategies of Gaussian-based Joint Rendering and Prior-guided Sparse Modeling. This occupancy-centric approach reduces the generation burden, especially for intricate scenes, while providing detailed intermediate representations for the subsequent generation stages. Extensive experiments demonstrate that UniScene outperforms previous SOTAs in the occupancy, video, and LiDAR generation, which also indeed benefits downstream driving tasks. The Project is available at https://arlo0o.github.io/uniscene/.

End-to-end autonomous driving unifies tasks in a differentiable framework, enabling planning-oriented optimization and attracting growing attention. Existing methods aggregate historical information either through dense historical bird's-eye-view (BEV) features or by querying a sparse memory bank, following paradigms inherited from detection. We argue that these paradigms either omit historical information in motion planning or fail to align with its multi-step nature, which requires predicting or planning multiple future time steps. In line with the philosophy of "future is a continuation of past", we propose BridgeAD, which reformulates motion and planning queries as multistep queries to differentiate the queries for each future time step. This design enables the effective use of historical prediction and planning by applying them to the appropriate parts of the end-to-end system based on the time steps, which improves both perception and motion planning. Specifically, historical queries for the current frame are combined with perception, while queries for future frames are integrated with motion planning. In this way, we bridge the gap between past and future by aggregating historical insights at every time step, enhancing the overall coherence and accuracy of the end-to-end autonomous driving pipeline. Extensive experiments on the nuScenes dataset in both open-loop and closed-loop settings demonstrate that BridgeAD achieves state-of-the-art performance.

Recent advances in diffusion models have greatly improved text-driven video generation. However, training models for long video generation demands significant computational power and extensive data, leading most video diffusion models to be limited to a small number of frames. Existing training-free methods that attempt to generate long videos using pre-trained short video diffusion models often struggle with issues such as insufficient motion dynamics and degraded video fidelity. In this paper, we present Brick-Diffusion, a novel, training-free approach capable of generating long videos of arbitrary length. Our method introduces a brick-to-wall denoising strategy, where the latent is denoised in segments, with a stride applied in subsequent iterations. This process mimics the construction of a staggered brick wall, where each brick represents a denoised segment, enabling communication between frames and improving overall video quality. Through quantitative and qualitative evaluations, we demonstrate that Brick-Diffusion outperforms existing baseline methods in generating high-fidelity videos.

The realistic reconstruction of street scenes is critical for developing real-world simulators in autonomous driving. Most existing methods rely on object pose annotations, using these poses to reconstruct dynamic objects and move them during the rendering process. This dependence on high-precision object annotations limits large-scale and extensive scene reconstruction. To address this challenge, we propose Bézier curve Gaussian splatting (BézierGS), which represents the motion trajectories of dynamic objects using learnable Bézier curves. This approach fully leverages the temporal information of dynamic objects and, through learnable curve modeling, automatically corrects pose errors. By introducing additional supervision on dynamic object rendering and inter-curve consistency constraints, we achieve reasonable and accurate separation and reconstruction of scene elements. Extensive experiments on the Waymo Open Dataset and the nuPlan benchmark demonstrate that BézierGS outperforms state-of-the-art alternatives in both dynamic and static scene components reconstruction and novel view synthesis.

LiDAR novel view synthesis (NVS) has emerged as a novel task within LiDAR simulation, offering valuable simulated point cloud data from novel viewpoints to aid in autonomous driving systems. However, existing LiDAR NVS methods typically rely on neural radiance fields (NeRF) as their 3D representation, which incurs significant computational costs in both training and rendering. Moreover, NeRF and its variants are designed for symmetrical scenes, making them ill-suited for driving scenarios. To address these challenges, we propose GS-LiDAR, a novel framework for generating realistic LiDAR point clouds with panoramic Gaussian splatting. Our approach employs 2D Gaussian primitives with periodic vibration properties, allowing for precise geometric reconstruction of both static and dynamic elements in driving scenarios. We further introduce a novel panoramic rendering technique with explicit ray-splat intersection, guided by panoramic LiDAR supervision. By incorporating intensity and ray-drop spherical harmonic (SH) coefficients into the Gaussian primitives, we enhance the realism of the rendered point clouds. Extensive experiments on KITTI-360 and nuScenes demonstrate the superiority of our method in terms of quantitative metrics, visual quality, as well as training and rendering efficiency.

Recent advancements in 3D generation are predominantly propelled by improvements in 3D-aware image diffusion models. These models are pretrained on Internet-scale image data and fine-tuned on massive 3D data, offering the capability of producing highly consistent multi-view images. However, due to the scarcity of synchronized multi-view video data, it remains challenging to adapt this paradigm to 4D generation directly. Despite that, the available video and 3D data are adequate for training video and multi-view diffusion models separately that can provide satisfactory dynamic and geometric priors respectively. To take advantage of both, this paper presents Diffusion$^2$, a novel framework for dynamic 3D content creation that reconciles the knowledge about geometric consistency and temporal smoothness from these models to directly sample dense multi-view multi-frame images which can be employed to optimize continuous 4D representation. Specifically, we design a simple yet effective denoising strategy via score composition of pretrained video and multi-view diffusion models based on the probability structure of the target image array. To alleviate the potential conflicts between two heterogeneous scores, we further introduce variance-reducing sampling via interpolated steps, facilitating smooth and stable generation. Owing to the high parallelism of the proposed image generation process and the efficiency of the modern 4D reconstruction pipeline, our framework can generate 4D content within few minutes. Notably, our method circumvents the reliance on expensive and hard-to-scale 4D data, thereby having the potential to benefit from the scaling of the foundation video and multi-view diffusion models. Extensive experiments demonstrate the efficacy of our proposed framework in generating highly seamless and consistent 4D assets under various types of conditions.

Novel view synthesis has experienced significant advancements owing to increasingly capable NeRF- and 3DGS-based methods. However, reflective object reconstruction remains challenging, lacking a proper solution to achieve real-time, high-quality rendering while accommodating inter-reflection. To fill this gap, we introduce a Reflective Gaussian splatting (Ref-Gaussian) framework characterized with two components: (I) Physically based deferred rendering that empowers the rendering equation with pixel-level material properties via formulating split-sum approximation; (II) Gaussian-grounded inter-reflection that realizes the desired inter-reflection function within a Gaussian splatting paradigm for the first time. To enhance geometry modeling, we further introduce material-aware normal propagation and an initial per-Gaussian shading stage, along with 2D Gaussian primitives. Extensive experiments on standard datasets demonstrate that Ref-Gaussian surpasses existing approaches in terms of quantitative metrics, visual quality, and compute efficiency. Further, we show that our method serves as a unified solution for both reflective and non-reflective scenes, going beyond the previous alternatives focusing on only reflective scenes. Also, we illustrate that Ref-Gaussian supports more applications such as relighting and editing.

Text-driven video generation has advanced significantly due to developments in diffusion models. Beyond the training and sampling phases, recent studies have investigated noise priors of diffusion models, as improved noise priors yield better generation results. One recent approach employs the Fourier transform to manipulate noise, marking the initial exploration of frequency operations in this context. However, it often generates videos that lack motion dynamics and imaging details. In this work, we provide a comprehensive theoretical analysis of the variance decay issue present in existing methods, contributing to the loss of details and motion dynamics. Recognizing the critical impact of noise distribution on generation quality, we introduce FreqPrior, a novel noise initialization strategy that refines noise in the frequency domain. Our method features a novel filtering technique designed to address different frequency signals while maintaining the noise prior distribution that closely approximates a standard Gaussian distribution. Additionally, we propose a partial sampling process by perturbing the latent at an intermediate timestep while finding the noise prior, significantly reducing inference time without compromising quality. Extensive experiments on VBench demonstrate that our method achieves the highest scores in both quality and semantic assessments, resulting in the best overall total score. These results highlight the superiority of our proposed noise prior.

Road surface reconstruction is crucial for autonomous driving, providing accurate and up-to-date road geometry for navigation, safety assessment, and infrastructure maintenance. Camera-based methods have become increasingly cost-effective and scalable for this task. However, achieving high-quality large-scale reconstruction remains challenging due to inconsistent observations of the same road surface points. These inconsistencies arise both within single sessions-caused by factors such as vehicle shadows and exposure shifts-and across multiple sessions, where changes in lighting conditions and viewpoints further exacerbate the problem. To address these challenges, we propose MS-Road, a camera-based approach for large-scale road surface reconstruction with strong geometric consistency. MS-Road leveraging self-supervised learning and multi-view consistent constrain to tackle two key issues: inconsistency in road appearance across different observations and inaccurate road height localization. By enforcing spatiotemporal consistency in both geometric and visual aspects, our method produces more reliable reconstructions within and across sessions. Experiments on two public datasets and a real-world dataset demonstrate that our approach achieves robust and high-fidelity reconstruction under diverse and challenging conditions.

Unified vision-language models have made significant progress in multimodal understanding and generation, yet they largely fall short in producing multimodal interleaved outputs, which is a crucial capability for tasks like visual storytelling and step-by-step visual reasoning. In this work, we propose a reinforcement learning-based post-training strategy to unlock this capability in existing unified models, without relying on large-scale multimodal interleaved datasets. We begin with a warm-up stage using a hybrid dataset comprising curated interleaved sequences and limited data for multimodal understanding and text-to-image generation, which exposes the model to interleaved generation patterns while preserving its pretrained capabilities. To further refine interleaved generation, we propose a unified policy optimization framework that extends Group Relative Policy Optimization (GRPO) to the multimodal setting. Our approach jointly models text and image generation within a single decoding trajectory and optimizes it with our novel hybrid rewards covering textual relevance, visual-text alignment, and structural fidelity. Additionally, we incorporate process-level rewards to provide step-wise guidance, enhancing training efficiency in complex multimodal tasks. Experiments on MMIE and InterleavedBench demonstrate that our approach significantly enhances the quality and coherence of multimodal interleaved generation.

Motion simulation, prediction and planning are foundational tasks in autonomous driving, each essential for modeling and reasoning about dynamic traffic scenarios. While often addressed in isolation due to their differing objectives, such as generating diverse motion states or estimating optimal trajectories, these tasks inherently depend on shared capabilities: understanding multi-agent interactions, modeling motion behaviors, and reasoning over temporal and spatial dynamics. Despite this underlying commonality, existing approaches typically adopt specialized model designs, which hinders cross-task generalization and system scalability. More critically, this separation overlooks the potential mutual benefits among tasks. Motivated by these observations, we propose **UniMotion**, a unified motion framework that captures shared structures across motion tasks while accommodating their individual requirements. Built on a decoder-only Transformer architecture, UniMotion employs dedicated interaction modes and tailored training strategies to simultaneously support these motion tasks. This unified design not only enables joint optimization and representation sharing but also allows for targeted fine-tuning to specialize in individual tasks when needed. Extensive experiments on the Waymo Open Motion Dataset (WOMD) demonstrate that joint training leads to robust generalization and effective task integration. With further fine-tuning, UniMotion achieves state-of-the-art performance across a range of motion tasks, establishing it as a versatile and scalable solution for autonomous driving.

Recent advances in LVLMs have improved vision-language understanding, but they still struggle with spatial perception, limiting their ability to reason about complex 3D scenes. Unlike previous approaches that incorporate 3D representations into models to improve spatial understanding, we aim to unlock the potential of VLMs by leveraging spatially relevant image data. To this end, we introduce a novel 2D spatial data generation and annotation pipeline built upon scene data with 3D ground-truth. This pipeline enables the creation of a diverse set of spatial tasks, ranging from basic perception tasks to more complex reasoning tasks. Leveraging this pipeline, we construct SPAR-7M, a large-scale dataset generated from thousands of scenes across multiple public datasets. In addition, we introduce SPAR-Bench, a benchmark designed to offer a more comprehensive evaluation of spatial capabilities compared to existing spatial benchmarks, supporting both single-view and multi-view inputs. Training on both SPAR-7M and large-scale 2D datasets enables our models to achieve state-of-the-art performance on 2D spatial benchmarks. Further fine-tuning on 3D task-specific datasets yields competitive results, underscoring the effectiveness of our dataset in enhancing spatial reasoning.

Leveraging diverse robotic data for pretraining remains a critical challenge. Existing methods typically model the dataset’s action distribution using simple observations as inputs. However, these inputs are often incomplete, resulting in a dispersed conditional action distribution—an issue we refer to as coordinate system chaos and state chaos. This inconsistency significantly hampers pretraining efficiency. To address this, we propose 4D-VLA, a novel approach that effectively integrates 4D information into the input to mitigate these sources of chaos. Our model introduces depth and temporal information into visual features with sequential RGB-D inputs, aligning the coordinate systems of the robot and the scene. This alignment endows the model with strong spatiotemporal reasoning capabilities while minimizing training overhead. Additionally, we introduce Memory bank sampling, a frame sampling strategy designed to extract informative frames from historical images, further improving effectiveness and efficiency. Experimental results demonstrate that our pretraining method and architectural components substantially enhance model performance. In both simulated and real-world experiments, our model achieves a significant increase in success rate over OpenVLA.To further assess spatial perception and generalization to novel views, we introduce MV-Bench, a multi-view simulation benchmark. Our model consistently outperforms existing methods, demonstrating stronger spatial understanding and adaptability.

End-to-end autonomous driving methods aim to directly map raw sensor inputs to future driving actions such as planned trajectories, bypassing traditional modular pipelines. While these approaches have shown promise, they often operate under a one-shot paradigm that relies heavily on the current scene context, potentially underestimating the importance of scene dynamics and their temporal evolution. This limitation restricts the model’s ability to make informed and adaptive decisions in complex driving scenarios. We propose a new perspective: the future trajectory of an autonomous vehicle is closely intertwined with the evolving dynamics of its environment, and conversely, the vehicle’s own future states can influence how the surrounding scene unfolds. Motivated by this bidirectional relationship, we introduce **SeerDrive**, a novel end-to-end framework that jointly models future scene evolution and trajectory planning in a closed-loop manner. Our method first predicts future bird’s-eye view (BEV) representations to anticipate the dynamics of the surrounding scene, then leverages this foresight to generate future-context-aware trajectories. Two key components enable this: (1) future-aware planning, which injects predicted BEV features into the trajectory planner, and (2) iterative scene modeling and vehicle planning, which refines both future scene prediction and trajectory generation through collaborative optimization. Extensive experiments on the NAVSIM and nuScenes benchmarks show that SeerDrive significantly outperforms existing state-of-the-art methods.

Autonomous driving simulation system plays a crucial role in enhancing self-driving data and simulating complex and rare traffic scenarios, ensuring navigation safety. However, traditional simulation systems, which often heavily rely on manual modeling and 2D image editing, struggled with scaling to extensive scenes and generating realistic simulation data. In this study, we present S-NeRF++, an innovative autonomous driving simulation system based on neural reconstruction. Trained on widely-used self-driving datasets, such as nuScenes and Waymo, S-NeRF++ can generate a large number of realistic street scenes and foreground objects with high rendering quality as well as offering considerable flexibility in manipulation and simulation. Specifically, S-NeRF++ is an enhanced neural radiance field for synthesizing large-scale scenes and moving vehicles, with improved scene parameterization and camera pose learning. The system effectively utilizes noisy and sparse LiDAR data to refine training and address depth outliers, ensuring high-quality reconstruction and novel-view rendering. It also provides a diverse foreground asset bank by reconstructing and generating different foreground vehicles to support comprehensive scenario creation. Moreover, we have developed an advanced foreground-background fusion pipeline that skillfully integrates illumination and shadow effects, further enhancing the realism of our simulations. With the high-quality simulated data provided by our S-NeRF++, we found the perception methods enjoy performance boosts on several autonomous driving downstream tasks, further demonstrating our proposed simulator's effectiveness.

Existing top-performance autonomous driving systems typically rely on the multi-modal fusion strategy for reliable scene understanding. This design is however fundamentally restricted due to overlooking the modality-specific strengths and finally hampering the model performance. To address this limitation, in this work, we introduce a novel modality interaction strategy that allows individual per-modality representations to be learned and maintained throughout, enabling their unique characteristics to be exploited during the whole perception pipeline. To demonstrate the effectiveness of the proposed strategy, we design DeepInteraction++, a multi-modal interaction framework characterized by a multi-modal representational interaction encoder and a multi-modal predictive interaction decoder. Specifically, the encoder is implemented as a dual-stream Transformer with specialized attention operation for information exchange and integration between separate modality-specific representations. Our multi-modal representational learning incorporates both object-centric, precise sampling-based feature alignment and global dense information spreading, essential for the more challenging planning task. The decoder is designed to iteratively refine the predictions by alternately aggregating information from separate representations in a unified modality-agnostic manner, realizing multi-modal predictive interaction. Extensive experiments demonstrate the superior performance of the proposed framework on both 3D object detection and end-to-end autonomous driving tasks.

Understanding road structures is crucial for autonomous driving. Intricate road structures are often depicted using lane graphs, which include centerline curves and connections forming a Directed Acyclic Graph (DAG). Accurate extraction of lane graphs relies on precisely estimating vertex and edge information within the DAG. Recent research highlights Transformer-based language models' impressive sequence prediction abilities, making them effective for learning graph representations when graph data are encoded as sequences. However, existing studies focus mainly on modeling vertices explicitly, leaving edge information simply embedded in the network. Consequently, these approaches fall short in the task of lane graph extraction. To address this, we introduce LaneGraph2Seq, a novel approach for lane graph extraction. It leverages a language model with vertex-edge encoding and connectivity enhancement. Our serialization strategy includes a vertex-centric depth-first traversal and a concise edge-based partition sequence. Additionally, we use classifier-free guidance combined with nucleus sampling to improve lane connectivity. We validate our method on prominent datasets, nuScenes and Argoverse 2, showcasing consistent and compelling results. Our LaneGraph2Seq approach demonstrates superior performance compared to state-of-the-art techniques in lane graph extraction.

In this paper, we present Consistent4D, a novel approach for generating 4D dynamic objects from uncalibrated monocular videos. Uniquely, we cast the 360-degree dynamic object reconstruction as a 4D generation problem, eliminating the need for tedious multi-view data collection and camera calibration. This is achieved by leveraging the object-level 3D-aware image diffusion model as the primary supervision signal for training dynamic Neural Radiance Fields (DyNeRF). Specifically, we propose a cascade DyNeRF to facilitate stable convergence and temporal continuity under the time-discrete supervision signal. To achieve spatial and temporal consistency of the 4D generation, an interpolation-driven consistency loss is further introduced, which aligns the rendered frames with the interpolated frames from a pre-trained video interpolation model. Extensive experiments show that the proposed Consistent4D significantly outperforms previous 4D reconstruction approaches as well as per-frame 3D generation approaches, opening up new possibilities for 4D dynamic object generation from a single-view uncalibrated video. Project page: https://consistent4d.github.io

High-resolution 3D object generation remains a challenging task primarily due to the limited availability of comprehensive annotated training data. Recent advancements have aimed to overcome this constraint by harnessing image generative models, pretrained on extensive curated web datasets, using knowledge transfer techniques like Score Distillation Sampling (SDS). Efficiently addressing the requirements of high-resolution rendering often necessitates the adoption of latent representation-based models, such as the Latent Diffusion Model (LDM). In this framework, a significant challenge arises: To compute gradients for individual image pixels, it is necessary to backpropagate gradients from the designated latent space through the frozen components of the image model, such as the VAE encoder used within LDM. However, this gradient propagation pathway has never been optimized, remaining uncontrolled during training. We find that the unregulated gradients adversely affect the 3D model's capacity in acquiring texture-related information from the image generative model, leading to poor quality appearance synthesis. To address this overarching challenge, we propose an innovative operation termed Pixel-wise Gradient Clipping (PGC) designed for seamless integration into existing 3D generative models, thereby enhancing their synthesis quality. Specifically, we control the magnitude of stochastic gradients by clipping the pixel-wise gradients efficiently, while preserving crucial texture-related gradient directions. Despite this simplicity and minimal extra cost, extensive experiments demonstrate the efficacy of our PGC in enhancing the performance of existing 3D generative models for high-resolution object rendering.

Reconstructing dynamic 3D scenes from 2D images and generating diverse views over time is challenging due to scene complexity and temporal dynamics. Despite advancements in neural implicit models, limitations persist: (i) Inadequate Scene Structure: Existing methods struggle to reveal the spatial and temporal structure of dynamic scenes from directly learning the complex 6D plenoptic function. (ii) Scaling Deformation Modeling: Explicitly modeling scene element deformation becomes impractical for complex dynamics. To address these issues, we consider the spacetime as an entirety and propose to approximate the underlying spatio-temporal 4D volume of a dynamic scene by optimizing a collection of 4D primitives, with explicit geometry and appearance modeling. Learning to optimize the 4D primitives enables us to synthesize novel views at any desired time with our tailored rendering routine. Our model is conceptually simple, consisting of a 4D Gaussian parameterized by anisotropic ellipses that can rotate arbitrarily in space and time, as well as view-dependent and time-evolved appearance represented by the coefficient of 4D spherindrical harmonics. This approach offers simplicity, flexibility for variable-length video and end-to-end training, and efficient real-time rendering, making it suitable for capturing complex dynamic scene motions. Experiments across various benchmarks, including monocular and multi-view scenarios, demonstrate our 4DGS model's superior visual quality and efficiency.

3D reconstruction has been widely used in autonomous navigation fields of mobile robotics. However, the former research can only provide the basic geometry structure without the capability of open-world scene understanding, limiting advanced tasks like human interaction and visual navigation. Moreover, traditional 3D scene understanding approaches rely on expensive labeled 3D datasets to train a model for a single task with supervision. Thus, geometric reconstruction with zero-shot scene understanding i.e. Open vocabulary 3D Understanding and Reconstruction, is crucial for the future development of mobile robots. In this paper, we propose OpenOcc, a novel framework unifying the 3D scene reconstruction and open vocabulary understanding with neural radiance fields. We model the geometric structure of the scene with occupancy representation and distill the pre-trained open vocabulary model into a 3D language field via volume rendering for zero-shot inference. Furthermore, a novel semantic-aware confidence propagation (SCP) method has been proposed to relieve the issue of language field representation degeneracy caused by inconsistent measurements in distilled features. Experimental results show that our approach achieves competitive performance in 3D scene understanding tasks, especially for small and long-tail objects.

3D representation is essential to the significant advance of 3D generation with 2D diffusion priors. As a flexible representation, NeRF has been first adopted for 3D representation. With density-based volumetric rendering, it however suffers both intensive computational overhead and inaccurate mesh extraction. Using a signed distance field and Marching Tetrahedra, DMTet allows for precise mesh extraction and real-time rendering but is limited in handling large topological changes in meshes, leading to optimization challenges. Alternatively, 3D Gaussian Splatting (3DGS) is favored in both training and rendering efficiency while falling short in mesh extraction. In this work, we introduce a novel 3D representation, Tetrahedron Splatting (TeT-Splatting), that supports easy convergence during optimization, precise mesh extraction, and real-time rendering simultaneously. This is achieved by integrating surface-based volumetric rendering within a structured tetrahedral grid while preserving the desired ability of precise mesh extraction, and a tile-based differentiable tetrahedron rasterizer. Furthermore, we incorporate eikonal and normal consistency regularization terms for the signed distance field to improve generation quality and stability. Critically, our representation can be trained without mesh extraction, making the optimization process easier to converge. Our TeT-Splatting can be readily integrated in existing 3D generation pipelines, along with polygonal mesh for texture optimization. Extensive experiments show that our TeT-Splatting strikes a superior tradeoff among convergence speed, render efficiency, and mesh quality as compared to previous alternatives under varying 3D generation settings.

Large language models (LLMs) have demonstrated exceptional capabilities in text understanding, which has paved the way for their expansion into video LLMs (Vid-LLMs) to analyze video data. However, current Vid-LLMs struggle to simultaneously retain high-quality frame-level semantic information (i.e., a sufficient number of tokens per frame) and comprehensive video-level temporal information (i.e., an adequate number of sampled frames per video). This limitation hinders the advancement of Vid-LLMs towards fine-grained video understanding. To address this issue, we introduce the SlowFocus mechanism, which significantly enhances the equivalent sampling frequency without compromising the quality of frame-level visual tokens. SlowFocus begins by identifying the query-related temporal segment based on the posed question, then performs dense sampling on this segment to extract local high-frequency features. A multi-frequency mixing attention module is further leveraged to aggregate these local high-frequency details with global low-frequency contexts for enhanced temporal comprehension. Additionally, to tailor Vid-LLMs to this innovative mechanism, we introduce a set of training strategies aimed at bolstering both temporal grounding and detailed temporal reasoning capabilities. Furthermore, we establish FineAction-CGR, a benchmark specifically devised to assess the ability of Vid-LLMs to process fine-grained temporal understanding tasks. Comprehensive experiments demonstrate the superiority of our mechanism across both existing public video understanding benchmarks and our proposed FineAction-CGR.

Motion forecasting for agents in autonomous driving is highly challenging due to the numerous possibilities for each agent's next action and their complex interactions in space and time. In real applications, motion forecasting takes place repeatedly and continuously as the self-driving car moves. However, existing forecasting methods typically process each driving scene within a certain range independently, totally ignoring the situational and contextual relationships between successive driving scenes. This significantly simplifies the forecasting task, making the solutions suboptimal and inefficient to use in practice. To address this fundamental limitation, we propose a novel motion forecasting framework for continuous driving, named RealMotion. It comprises two integral streams both at the scene level: (1) The scene context stream progressively accumulates historical scene information until the present moment, capturing temporal interactive relationships among scene elements. (2) The agent trajectory stream optimizes current forecasting by sequentially relaying past predictions. Besides, a data reorganization strategy is introduced to narrow the gap between existing benchmarks and real-world applications, consistent with our network. These approaches enable exploiting more broadly the situational and progressive insights of dynamic motion across space and time. Extensive experiments on Argoverse series with different settings demonstrate that our RealMotion achieves state-of-the-art performance, along with the advantage of efficient real-world inference.

Achieving robust and precise pose estimation in dynamic scenes is a significant research challenge in Visual Simultaneous Localization and Mapping (SLAM). Recent advancements integrating Gaussian Splatting into SLAM systems have proven effective in creating high-quality renderings using explicit 3D Gaussian models, significantly improving environmental reconstruction fidelity. However, these approaches depend on a static environment assumption and face challenges in dynamic environments due to inconsistent observations of geometry and photometry. To address this problem, we propose DG-SLAM, the first robust dynamic visual SLAM system grounded in 3D Gaussians, which provides precise camera pose estimation alongside high-fidelity reconstructions. Specifically, we propose effective strategies, including motion mask generation, adaptive Gaussian point management, and a hybrid camera tracking algorithm to improve the accuracy and robustness of pose estimation. Extensive experiments demonstrate that DG-SLAM delivers state-of-the-art performance in camera pose estimation, map reconstruction, and novel-view synthesis in dynamic scenes, outperforming existing methods meanwhile preserving real-time rendering ability.

Accurate motion forecasting for traffic agents is crucial for ensuring the safety and efficiency of autonomous driving systems in dynamically changing environments. Mainstream methods adopt a one-query-one-trajectory paradigm, where each query corresponds to a unique trajectory for predicting multi-modal trajectories. While straightforward and effective, the absence of detailed representation of future trajectories may yield suboptimal outcomes, given that the agent states dynamically evolve over time. To address this problem, we introduce DeMo, a framework that decouples multi-modal trajectory queries into two types: mode queries capturing distinct directional intentions and state queries tracking the agent's dynamic states over time. By leveraging this format, we separately optimize the multi-modality and dynamic evolutionary properties of trajectories. Subsequently, the mode and state queries are integrated to obtain a comprehensive and detailed representation of the trajectories. To achieve these operations, we additionally introduce combined Attention and Mamba techniques for global information aggregation and state sequence modeling, leveraging their respective strengths. Extensive experiments on both the Argoverse 2 and nuScenes benchmarks demonstrate that our DeMo achieves state-of-the-art performance in motion forecasting. In addition, we will make our code and models publicly available.

3D object detection in autonomous driving aims to reason “what” and “where” the objects of interest present in a 3D world. Following the conventional wisdom of previous 2D object detection, existing methods often adopt the canonical Cartesian coordinate system with perpendicular axis. However, we conjugate that this does not fit the nature of the ego car’s perspective, as each onboard camera perceives the world in shape of wedge intrinsic to the imaging geometry with radical (non perpendicular) axis. Hence, in this paper we advocate the exploitation of the Polar coordinate system and propose a new Polar Transformer (PolarFormer) for more accurate 3D object detection in the bird’s-eye-view (BEV) taking as input only multi-camera 2D images. Specifically, we design a cross-attention based Polar detection head without restriction to the shape of input structure to deal with irregular Polar grids. For tackling the unconstrained object scale variations along Polar’s distance dimension, we further introduce a multi-scale Polar representation learning strategy. As a result, our model can make best use of the Polar representation rasterized via attending to the corresponding image observation in a sequence-to-sequence fashion subject to the geometric constraints. Thorough experiments on the nuScenes dataset demonstrate that our PolarFormer outperforms significantly state-of-the-art 3D object detection alternatives.

We present Generative Semantic Segmentation (GSS), a generative learning approach for semantic segmentation. Uniquely, we cast semantic segmentation as an image-conditioned mask generation problem. This is achieved by replacing the conventional per-pixel discriminative learning with a latent prior learning process. Specifically, we model the variational posterior distribution of latent vari-ables given the segmentation mask. To that end, the segmentation mask is expressed with a special type of image (dubbed as maskige). This posterior distribution allows to generate segmentation masks unconditionally. To achieve semantic segmentation on a given image, we further intro-duce a conditioning network. It is optimized by minimizing the divergence between the posterior distribution of maskige (i.e. segmentation masks) and the latent prior distribution of input training images. Extensive experiments on standard benchmarks show that our GSS can perform competitively to prior art alternatives in the standard semantic segmentation setting, whilst achieving a new state of the art in the more challenging cross-domain setting.

In the fashion domain, there exists a variety of vision-and-language (V+L) tasks, including cross-modal retrieval, text-guided image retrieval, multi-modal classification, and image captioning. They differ drastically in each individual input/output format and dataset size. It has been common to design a task-specific model and fine-tune it independently from a pre-trained V+l model (e.g., CLIP). This results in parameter inefficiency and inability to exploit inter-task relatedness. To address such issues, we propose a novel FAshion-focused Multi-task Efficient learning method for Vision-and-Language tasks (FAME-ViL) in this work. Compared with existing approaches, FAME-ViL applies a single model for multiple heterogeneous fashion tasks, therefore being much more parameter-efficient. It is enabled by two novel components: (1) a task-versatile architecture with cross-attention adapters and task-specific adapters integrated into a unified V+L model, and (2) a stable and effective multi-task training strategy that supports learning from heterogeneous data and prevents negative transfer. Extensive experiments on four fashion tasks show that our FAME-ViL can save 61.5% of parameters over alternatives, while significantly outperforming the conventional independently trained single-task models. Code is available at https://github.com/BrandonHanx/FAME-ViL.

The extraction of road network is essential for the generation of high-definition maps since it enables the precise localization of road landmarks and their interconnections. However, generating road network poses a significant challenge due to the conflicting underlying combination of Euclidean (e.g., road landmarks location) and non-Euclidean (e.g., road topological connectivity) structures. Existing methods struggle to merge the two types of data domains effectively, but few of them address it properly. Instead, our work establishes a unified representation of both types of data domain by projecting both Euclidean and non-Euclidean data into an integer series called RoadNet Sequence. Further than modeling an auto-regressive sequence-to-sequence Transformer model to understand RoadNet Sequence, we decouple the dependency of RoadNet Sequence into a mixture of auto-regressive and non-autoregressive dependency. Building on this, our proposed non-autoregressive sequence-to-sequence approach leverages non-autoregressive dependencies while fixing the gap towards auto-regressive dependencies, resulting in success on both efficiency and accuracy. Extensive experiments on nuScenes dataset demonstrate the superiority of Road-Net Sequence representation and the non-autoregressive approach compared to existing state-of-the-art alternatives.

Recently, polar-based representation has shown promising properties in perceptual tasks. In addition to Cartesian-based approaches, which separate point clouds unevenly, representing point clouds as polar grids has been recognized as an alternative due to (1) its advantage in robust performance under different resolutions and (2) its superiority in streaming-based approaches. However, state-of-the-art polar-based detection methods inevitably suffer from the feature distortion problem because of the non-uniform division of polar representation, resulting in a non-negligible performance gap compared to Cartesian-based approaches. To tackle this issue, we present PARTNER, a novel 3D object detector in the polar coordinate. PARTNER alleviates the dilemma of feature distortion with global representation re-alignment and facilitates the regression by introducing instance-level geometric information into the detection head. Extensive experiments show overwhelming advantages in streaming-based detection and different resolutions. Furthermore, our method outperforms the previous polar-based works with remarkable margins of 3.68% and 9.15% on Waymo and ONCE validation set, thus achieving competitive results over the state-of-the-art methods.

3D object detection from LiDAR point cloud is of critical importance for autonomous driving and robotics. While sequential point cloud has the potential to enhance 3D perception through temporal information, utilizing these temporal features effectively and efficiently remains a challenging problem. Based on the observation that the foreground information is sparsely distributed in LiDAR scenes, we believe sufficient knowledge can be provided by sparse format rather than dense maps. To this end, we propose to learn Significance-gUided Information for 3D Temporal detection (SUIT), which simplifies temporal information as sparse features for information fusion across frames. Specifically, we first introduce a significant sampling mechanism that extracts information-rich yet sparse features based on predicted object centroids. On top of that, we present an explicit geometric transformation learning technique, which learns the object-centric transformations among sparse features across frames. We evaluate our method on large-scale nuScenes and Waymo dataset, where our SUIT not only significantly reduces the memory and computation cost of temporal fusion, but also performs well over the state-of-the-art baselines.

Conventional 3D object detection approaches concentrate on bounding boxes representation learning with several parameters, i.e., localization, dimension, and orientation. Despite its popularity and universality, such a straightforward paradigm is sensitive to slight numerical deviations, especially in localization. By exploiting the property that point clouds are naturally captured on the surface of objects along with accurate location and intensity information, we introduce a new perspective that views bounding box regression as an implicit function. This leads to our proposed framework, termed Implicit Detection or ImpDet, which leverages implicit field learning for 3D object detection. Our ImpDet assigns specific values to points in different local 3D spaces, thereby high-quality boundaries can be generated by classifying points inside or outside the boundary. To solve the problem of sparsity on the object surface, we further present a simple yet efficient virtual sampling strategy to not only fill the empty region, but also learn rich semantic features to help refine the boundaries. Extensive experimental results on KITTI and Waymo benchmarks demonstrate the effectiveness and robustness of unifying implicit fields into object detection.

In this article, we introduce SiamMask, a framework to perform both visual object tracking and video object segmentation, in real-time, with the same simple method. We improve the offline training procedure of popular fully-convolutional Siamese approaches by augmenting their losses with a binary segmentation task. Once the offline training is completed, SiamMask only requires a single bounding box for initialization and can simultaneously carry out visual object tracking and segmentation at high frame-rates. Moreover, we show that it is possible to extend the framework to handle multiple object tracking and segmentation by simply re-using the multi-task model in a cascaded fashion. Experimental results show that our approach has high processing efficiency, at around 55 frames per second. It yields real-time state-of-the art results on visual-object tracking benchmarks, while at the same time demonstrating competitive performance at a high speed for video object segmentation benchmarks.

Modelling long-range dependencies is critical for scene understanding tasks in computer vision. Although convolution neural networks (CNNs) have excelled in many vision tasks, they are still limited in capturing long-range structured relationships as they typically consist of layers of local kernels. A fully-connected graph, such as the self-attention operation in Transformers, is beneficial for such modelling, however, its computational overhead is prohibitive. In this paper, we propose a dynamic graph message passing network, that significantly reduces the computational complexity compared to related works modelling a fully-connected graph. This is achieved by adaptively sampling nodes in the graph, conditioned on the input, for message passing. Based on the sampled nodes, we dynamically predict node-dependent filter weights and the affinity matrix for propagating information between them. This formulation allows us to design a self-attention module, and more importantly a new Transformer-based backbone network, that we use for both image classification pretraining, and for addressing various downstream tasks (e.g. object detection, instance and semantic segmentation). Using this model, we show significant improvements with respect to strong, state-of-the-art baselines on four different tasks. Our approach also outperforms fully-connected graphs while using substantially fewer floating-point operations and parameters. Code and models will be made publicly available at https://github.com/fudan-zvg/DGMN2.

In recent intelligent-robot-assisted surgery studies, an urgent issue is how to detect the motion of instruments and soft tissue accurately from intra-operative images. Although optical flow technology from computer vision is a powerful solution to the motion-tracking problem, it has difficulty obtaining the pixel-wise optical flow ground truth of real surgery videos for supervised learning. Thus, unsupervised learning methods are critical. However, current unsupervised methods face the challenge of heavy occlusion in the surgical scene. This paper proposes a novel unsupervised learning framework to estimate the motion from surgical images under occlusion. The framework consists of a Motion Decoupling Network to estimate the tissue and the instrument motion with different constraints. Notably, the network integrates a segmentation subnet that estimates the segmentation map of instruments in an unsupervised manner to obtain the occlusion region and improve the dual motion estimation. Additionally, a hybrid self-supervised strategy with occlusion completion is introduced to recover realistic vision clues. Extensive experiments on two surgical datasets show that the proposed method achieves accurate motion estimation for intra-operative scenes and outperforms other unsupervised methods, with a margin of 15% in accuracy. The average estimation error for tissue is less than 2.2 pixels on average for both surgical datasets.

Arguably the most representative application of artificial intelligence, autonomous driving systems usually rely on computer vision techniques to detect the situations of the external environment. Object detection underpins the ability of scene understanding in such systems. However, existing object detection algorithms often behave as a black box, so when a model fails, no information is available on When, Where and How the failure happened. In this paper, we propose a visual analytics approach to help model developers interpret the model failures. The system includes the micro- and macro-interpreting modules to address the interpretability problem of object detection in autonomous driving. The micro-interpreting module extracts and visualizes the features of a convolutional neural network (CNN) algorithm with density maps, while the macro-interpreting module provides spatial-temporal information of an autonomous driving vehicle and its environment. With the situation awareness of the spatial, temporal and neural network information, our system facilitates the understanding of the results of object detection algorithms, and helps the model developers better understand, tune and develop the models. We use real-world autonomous driving data to perform case studies by involving domain experts in computer vision and autonomous driving to evaluate our system. The results from our interviews with them show the effectiveness of our approach.

We present ONCE-3DLanes, a real-world autonomous driving dataset with lane layout annotation in 3D space. Conventional 2D lane detection from a monocular image yields poor performance of following planning and control tasks in autonomous driving due to the case of uneven road. Predicting the 3D lane layout is thus necessary and enables effective and safe driving. However, existing 3D lane detection datasets are either unpublished or synthesized from a simulated environment, severely hampering the development of this field. In this paper, we take steps towards addressing these issues. By exploiting the explicit relationship between point clouds and image pixels, a dataset annotation pipeline is designed to automatically generate high-quality 3D lane locations from 2D lane annotations in 211K road scenes. In addition, we present an extrinsic-free, anchorfree method, called SALAD, regressing the 3D coordinates of lanes in image view without converting the feature map into the bird's-eye view (BEV). To facilitate future research on 3D lane detection, we benchmark the dataset and provide a novel evaluation metric, performing extensive experiments of both existing approaches and our proposed method. The aim of our work is to revive the interest of 3D lane detection in a real-world scenario. We believe our work can lead to the expected and unexpected innovations in both academia and industry.

Existing top-performance 3D object detectors typically rely on the multi-modal fusion strategy. This design is however fundamentally restricted due to overlooking the modality-specific useful information and finally hampering the model performance. To address this limitation, in this work we introduce a novel modality interaction strategy where individual per-modality representations are learned and maintained throughout for enabling their unique characteristics to be exploited during object detection. To realize this proposed strategy, we design a DeepInteraction architecture characterized by a multi-modal representational interaction encoder and a multi-modal predictive interaction decoder. Experiments on the large-scale nuScenes dataset show that our proposed method surpasses all prior arts often by a large margin. Crucially, our method is ranked at the first position at the highly competitive nuScenes object detection leaderboard.

Deep learning based models have excelled in many computer vision tasks and appear to surpass humans' performance. However, these models require an avalanche of expensive human labeled training data and many iterations to train their large number of parameters. This severely limits their scalability to the real-world long-tail distributed categories, some of which are with a large number of instances, but with only a few manually annotated. Learning from such extremely limited labeled examples is known as Few-Shot Learning (FSL). Different to prior arts that leverage meta-learning or data augmentation strategies to alleviate this extremely data-scarce problem, this paper presents a statistical approach, dubbed Instance Credibility Inference (ICI) to exploit the support of unlabeled instances for few-shot visual recognition. Typically, we repurpose the self-taught learning paradigm to predict pseudo-labels of unlabeled instances with an initial classifier trained from the few shot and then select the most confident ones to augment the training set to re-train the classifier. This is achieved by constructing a (Generalized) Linear Model (LM/GLM) with incidental parameters to model the mapping from (un-)labeled features to their (pseudo-)labels, in which the sparsity of the incidental parameters indicates the credibility of the corresponding pseudo-labeled instance. We rank the credibility of pseudo-labeled instances along the regularization path of their corresponding incidental parameters, and the most trustworthy pseudo-labeled examples are preserved as the augmented labeled instances. This process is repeated until all the unlabeled samples are included in the expanded training set. Theoretically, under the conditions of restricted eigenvalue, irrepresentability, and large error, our approach is guaranteed to collect all the correctly-predicted pseudo-labeled instances from the noisy pseudo-labeled set. Extensive experiments under two few-shot settings show the effectiveness of our approach on four widely used few-shot visual recognition benchmark datasets including miniImageNet, tieredImageNet, CIFAR-FS, and CUB. Code and models are released at https://github.com/Yikai-Wang/ICI-FSL.

Unsupervised single image restoration approach, Deep Image Prior (DIP), aims to restore images by learning enough raw image statistic priors from the corrupted observation. However, it is not uncommon that an image is contaminated by the multiple unknown distortions. Thus it is hard to disentangle the clean and the hybrid distortion signals by solely relying on image prior learning to restore the images. To overcome this problem, we propose the Dual Prior Learning (DPL) method by taking both image and distortion priors into account. DPL goes beyond DIP by considering an additional step to explicitly learn the blended distortion prior. Furthermore, to coordinate the learning of two priors and avoid them learning the same knowledge, we exploit unpaired training data to enforce a weakly supervision in an adversarial manner to encourage disentangling two priors. Extensive experiments show the effectiveness and appealing performance of the proposed DPL on restoring images with challenging unknown blended distortions.

Few-shot learning (FSL) aims to recognize target classes by adapting the prior knowledge learned from source classes. Such knowledge usually resides in a deep embedding model for a general matching purpose of the support and query image pairs. The objective of this paper is to repurpose the contrastive learning for such matching to learn a few-shot embedding model. We make the following contributions: (i) We investigate the contrastive learning with Noise Contrastive Estimation (NCE) in a supervised manner for training a few-shot embedding model; (ii) We propose a novel contrastive training scheme dubbed infoPatch, exploiting the patch-wise relationship to substantially improve the popular infoNCE; (iii) We show that the embedding learned by the proposed infoPatch is more effective; (iv) Our model is thoroughly evaluated on few-shot recognition task; and demonstrates state-of-the-art results on miniImageNet and appealing performance on tieredImageNet, Fewshot-CIFAR100 (FC-100).

Few-shot learning (FSL), which aims to recognise new classes by adapting the learned knowledge with extremely limited few-shot (support) examples, remains an important open problem in computer vision. Most of the existing methods for feature alignment in few-shot learning only consider image-level or spatial-level alignment while omitting the channel disparity. Our insight is that these methods would lead to poor adaptation with redundant matching, and leveraging channel-wise adjustment is the key to well adapting the learned knowledge to new classes. Therefore, in this paper, we propose to learn a dynamic alignment, which can effectively highlight both query regions and channels according to different local support information. Specifically, this is achieved by first dynamically sampling the neighbourhood of the feature position conditioned on the input few shot, based on which we further predict a both position-dependent and channel-dependent Dynamic Meta-filter. The filter is used to align the query feature with position-specific and channel-specific knowledge. Moreover, we adopt Neural Ordinary Differential Equation (ODE) to enable a more accurate control of the alignment. In such a sense our model is able to better capture fine-grained semantic context of the few-shot example and thus facilitates dynamical knowledge adaptation for few-shot learning. The resulting framework establishes the new state-of-the-arts on major few-shot visual recognition benchmarks, including miniImageNet and tieredImageNet.

The objective of this paper is to learn context- and depth- aware feature representation to solve the problem of monocular 3D object detection. We make following contributions: (i) rather than appealing to the complicated pseudo-LiDAR based approach, we propose a depth-conditioned dynamic message propagation (DDMP) network to effectively integrate the multi-scale depth information with the image context; (ii) this is achieved by first adaptively sampling context-aware nodes in the image context and then dynamically predicting hybrid depth-dependent filter weights and affinity matrices for propagating information; (Hi) by augmenting a center-aware depth encoding (CDE) task, our method successfully alleviates the inaccurate depth prior; (iv) we thoroughly demonstrate the effectiveness of our proposed approach and show state-of-the-art results among the monocular-based approaches on the KITTI benchmark dataset. Particularly, we rank 1stin the highly competitive KITTI monocular 3D object detection track on the submission day (November 16th, 2020). Code and models are released at https: //github.com/fudan-zvg/DDMP

Deep models have shown their vulnerability when processing adversarial samples. As for the black-box attack, without access to the architecture and weights of the attacked model, training a substitute model for adversarial attacks has attracted wide attention. Previous substitute training approaches focus on stealing the knowledge of the target model based on real training data or synthetic data, without exploring what kind of data can further improve the transferability between the substitute and target models. In this paper, we propose a novel perspective substitute training that focuses on designing the distribution of data used in the knowledge stealing process. More specifically, a diverse data generation module is proposed to synthesize large-scale data with wide distribution. And adversarial substitute training strategy is introduced to focus on the data distributed near the decision boundary. The combination of these two modules can further boost the consistency of the substitute model and target model, which greatly improves the effectiveness of adversarial attack. Extensive experiments demonstrate the efficacy of our method against state-of-the-art competitors under non-target and target at-tack settings. Detailed visualization and analysis are also provided to help understand the advantage of our method.

Most recent semantic segmentation methods adopt a fully-convolutional network (FCN) with an encoder-decoder architecture. The encoder progressively reduces the spatial resolution and learns more abstract/semantic visual concepts with larger receptive fields. Since context modeling is critical for segmentation, the latest efforts have been focused on increasing the receptive field, through either dilated/atrous convolutions or inserting attention modules. However, the encoder-decoder based FCN architecture remains unchanged. In this paper, we aim to provide an alternative perspective by treating semantic segmentation as a sequence-to-sequence prediction task. Specifically, we deploy a pure transformer (i.e., without convolution and resolution reduction) to encode an image as a sequence of patches. With the global context modeled in every layer of the transformer, this encoder can be combined with a simple decoder to provide a powerful segmentation model, termed SEgmentation TRansformer (SETR). Extensive experiments show that SETR achieves new state of the art on ADE20K (50.28% mIoU), Pascal Context (55.83% mIoU) and competitive results on Cityscapes. Particularly, we achieve the first position in the highly competitive ADE20K test server leaderboard on the day of submission.

A few-shot semantic segmentation model is typically composed of a CNN encoder, a CNN decoder and a simple classifier (separating foreground and background pixels). Most existing methods meta-learn all three model components for fast adaptation to a new class. However, given that as few as a single support set image is available, effective model adaption of all three components to the new class is extremely challenging. In this work we propose to simplify the meta-learning task by focusing solely on the simplest component – the classifier, whilst leaving the en-coder and decoder to pre-training. We hypothesize that if we pretrain an off-the-shelf segmentation model over a set of diverse training classes with sufficient annotations, the encoder and decoder can capture rich discriminative features applicable for any unseen classes, rendering the sub-sequent meta-learning stage unnecessary. For the classifier meta-learning, we introduce a Classifier Weight Transformer (CWT) designed to dynamically adapt the support-set trained classifier’s weights to each query image in an inductive way. Extensive experiments on two standard bench-marks show that despite its simplicity, our method outperforms the state-of-the-art alternatives, often by a large margin. Code is available on https://github.com/zhiheLu/CWT-for-FSS.

Many video analysis tasks require temporal localization for the detection of content changes. However, most existing models developed for these tasks are pre-trained on general video action classification tasks. This is due to large scale annotation of temporal boundaries in untrimmed videos being expensive. Therefore, no suitable datasets exist that enable pre-training in a manner sensitive to temporal boundaries. In this paper for the first time, we investigate model pre-training for temporal localization by introducing a novel boundary-sensitive pretext (BSP) task. Instead of relying on costly manual annotations of temporal boundaries, we propose to synthesize temporal boundaries in existing video action classification datasets. By defining different ways of synthesizing boundaries, BSP can then be simply conducted in a self-supervised manner via the classification of the boundary types. This enables the learning of video representations that are much more transferable to downstream temporal localization tasks. Extensive experiments show that the proposed BSP is superior and complementary to the existing action classification-based pre-training counterpart, and achieves new state-of-the-art performance on several temporal localization tasks. Please visit our website for more details https://frostinassiky.github.io/bsp.

Low-cost monocular 3D object detection plays a fundamental role in autonomous driving, whereas its accuracy is still far from satisfactory. In this paper, we dig into the 3D object detection task and reformulate it as the sub-tasks of object localization and appearance perception, which benefits to a deep excavation of reciprocal information underlying the entire task. We introduce a Dynamic Feature Reflecting Network, named DFR-Net, which contains two novel standalone modules: (i) the Appearance-Localization Feature Reflecting module (ALFR) that first separates task-specific features and then self-mutually reflects the reciprocal features; (ii) the Dynamic Intra-Trading module (DIT) that adaptively realigns the training processes of various sub-tasks via a self-learning manner. Extensive experiments on the challenging KITTI dataset demonstrate the effectiveness and generalization of DFR-Net. We rank 1stamong all the monocular 3D object detectors in the KITTI test set (till March 16th, 2021). The proposed method is also easy to be plug-and-play in many cutting-edge 3D detection frameworks at negligible cost to boost performance. The code will be made publicly available.

Recent deep learning methods rely on a large amount of labeled data to achieve high performance. These methods may be impractical in some scenarios, where manual data annotation is costly or the samples of certain categories are scarce (e.g., tumor lesions, endangered animals and rare individual activities). When only limited annotated samples are available, these methods usually suffer from the overfitting problem severely, which degrades the performance significantly. In contrast, humans can recognize the objects in the images rapidly and correctly with their prior knowledge after exposed to only a few annotated samples. To simulate the learning schema of humans and relieve the reliance on the large-scale annotation benchmarks, researchers start shifting towards the few-shot learning problem: they try to learn a model to correctly recognize novel categories with only a few annotated samples.

Vision transformers (ViTs) have pushed the state-of-the-art for various visual recognition tasks by patch-wise image tokenization followed by self-attention. However, the employment of self-attention modules results in a quadratic complexity in both computation and memory usage. Various attempts on approximating the self-attention computation with linear complexity have been made in Natural Language Processing. However, an in-depth analysis in this work shows that they are either theoretically flawed or empirically ineffective for visual recognition. We further identify that their limitations are rooted in keeping the softmax self-attention during approximations. Specifically, conventional self-attention is computed by normalizing the scaled dot-product between token feature vectors. Keeping this softmax operation challenges any subsequent linearization efforts. Based on this insight, for the first time, a softmax-free transformer or SOFT is proposed. To remove softmax in self-attention, Gaussian kernel function is used to replace the dot-product similarity without further normalization. This enables a full self-attention matrix to be approximated via a low-rank matrix decomposition. The robustness of the approximation is achieved by calculating its Moore-Penrose inverse using a Newton-Raphson method. Extensive experiments on ImageNet show that our SOFT significantly improves the computational efficiency of existing ViT variants. Crucially, with a linear complexity, much longer token sequences are permitted in SOFT, resulting in superior trade-off between accuracy and complexity.

Recognizing and localizing objects in the 3D space is a crucial ability for an AI agent to perceive its surrounding environment. While significant progress has been achieved with expensive LiDAR point clouds, it poses a great challenge for 3D object detection given only a monocular image. While there exist different alternatives for tackling this problem, it is found that they are either equipped with heavy networks to fuse RGB and depth information or empirically ineffective to process millions of pseudo-LiDAR points. With in-depth examination, we realize that these limitations are rooted in inaccurate object localization. In this paper, we propose a novel and lightweight approach, dubbed {\em Progressive Coordinate Transforms} (PCT) to facilitate learning coordinate representations. Specifically, a localization boosting mechanism with confidence-aware loss is introduced to progressively refine the localization prediction. In addition, semantic image representation is also exploited to compensate for the usage of patch proposals. Despite being lightweight and simple, our strategy allows us to establish a new state-of-the-art among the monocular 3D detectors on the competitive KITTI benchmark. At the same time, our proposed PCT shows great generalization to most coordinate-based 3D detection frameworks.

Graph-based convolutional model such as non-local block has shown to be effective for strengthening the context modeling ability in convolutional neural networks (CNNs). However, its pixel-wise computational overhead is prohibitive which renders it unsuitable for high resolution imagery. In this paper, we explore the efficiency of context graph reasoning and propose a novel framework called Squeeze Reasoning. Instead of propagating information on the spatial map, we first learn to squeeze the input feature into a channel-wise global vector and perform reasoning within the single vector where the computation cost can be significantly reduced. Specifically, we build the node graph in the vector where each node represents an abstract semantic concept. The refined feature within the same semantic category results to be consistent, which is thus beneficial for downstream tasks. We show that our approach can be modularized as an end-to-end trained block and can be easily plugged into existing networks. Despite its simplicity and being lightweight, the proposed strategy allows us to establish the considerable results on different semantic segmentation datasets and shows significant improvements with respect to strong baselines on various other scene understanding tasks including object detection, instance segmentation and panoptic segmentation. Code is available at https://github.com/lxtGH/SFSegNets.

Modelling long-range contextual relationships is critical for pixel-wise prediction tasks such as semantic segmentation. However, convolutional neural networks (CNNs) are inherently limited to model such dependencies due to the naive structure in its building modules (e.g., local convolution kernel). While recent global aggregation methods are beneficial for long-range structure information modelling, they would oversmooth and bring noise to the regions contain fine details (e.g., boundaries and small objects), which are very much cared in the semantic segmentation task. To alleviate this problem, we propose to explore the local context for making the aggregated long-range relationship being distributed more accurately in local regions. In particular, we design a novel local distribution module which models the affinity map between global and local relationship for each pixel adaptively. Integrating existing global aggregation modules, we show that our approach can be modularized as an end-to-end trainable block and easily plugged into existing semantic segmentation networks, giving rise to the GALD networks. Despite its simplicity and versatility, our approach allows us to build new state of the art on major semantic segmentation benchmarks including Cityscapes, ADE20K, Pascal Context, Camvid and COCO-stuff. Code and trained models are released at https://github.com/lxtGH/GALD-DGCNet to foster further research.

Existing fully-supervised person re-identification (ReID) methods usually suffer from poor generalization capability caused by domain gaps. The key to solving this problem lies in filtering out identity-irrelevant interference and learning domain-invariant person representations. In this paper, we aim to design a generalizable person ReID framework which trains a model on source domains yet is able to generalize/perform well on target domains. To achieve this goal, we propose a simple yet effective Style Normalization and Restitution (SNR) module. Specifically, we filter out style variations (e.g., illumination, color contrast) by Instance Normalization (IN). However, such a process inevitably removes discriminative information. We propose to distill identity-relevant feature from the removed information and restitute it to the network to ensure high discrimination. For better disentanglement, we enforce a dual causal loss constraint in SNR to encourage the separation of identity-relevant features and identity-irrelevant features. Extensive experiments demonstrate the strong generalization capability of our framework. Our models empowered by the SNR modules significantly outperform the state-of-the-art domain generalization approaches on multiple widely-used person ReID benchmarks, and also show superiority on unsupervised domain adaptation.

Few-shot learning (FSL) aims to recognize new objects with extremely limited training data for each category. Previous efforts are made by either leveraging meta-learning paradigm or novel principles in data augmentation to alleviate this extremely data-scarce problem. In contrast, this paper presents a simple statistical approach, dubbed Instance Credibility Inference (ICI) to exploit the distribution support of unlabeled instances for few-shot learning. Specifically, we first train a linear classifier with the labeled few-shot examples and use it to infer the pseudo-labels for the unlabeled data. To measure the credibility of each pseudo-labeled instance, we then propose to solve another linear regression hypothesis by increasing the sparsity of the incidental parameters and rank the pseudo-labeled instances with their sparsity degree. We select the most trustworthy pseudo-labeled instances alongside the labeled examples to re-train the linear classifier. This process is iterated until all the unlabeled samples are included in the expanded training set, i.e. the pseudo-label is converged for unlabeled data pool. Extensive experiments under two few-shot settings show that our simple approach can establish new state-of-the-arts on four widely used few-shot learning benchmark datasets including miniImageNet, tieredImageNet, CIFAR-FS, and CUB. Our code is available at: https://github.com/Yikai-Wang/ICI-FSL

Modelling long-range dependencies is critical for scene understanding tasks in computer vision. Although CNNs have excelled in many vision tasks, they are still limited in capturing long-range structured relationships as they typically consist of layers of local kernels. A fully-connected graph is beneficial for such modelling, however, its computational overhead is prohibitive. We propose a dynamic graph message passing network, that significantly reduces the computational complexity compared to related works modelling a fully-connected graph. This is achieved by adaptively sampling nodes in the graph, conditioned on the input, for message passing. Based on the sampled nodes, we dynamically predict node-dependent filter weights and the affinity matrix for propagating information between them. Using this model, we show significant improvements with respect to strong, state-of-the-art baselines on three different tasks and backbone architectures. Our approach also outperforms fully-connected graphs while using substantially fewer floating-point operations and parameters.

Humans can easily recognize actions with only a few examples given, while the existing video recognition models still heavily rely on the large-scale labeled data inputs. This observation has motivated an increasing interest in few-shot video action recognition, which aims at learning new actions with only very few labeled samples. In this paper, we propose a depth guided Adaptive Meta-Fusion Network for few-shot video recognition which is termed as AMeFu-Net. Concretely, we tackle the few-shot recognition problem from three aspects: firstly, we alleviate this extremely data-scarce problem by introducing depth information as a carrier of the scene, which will bring extra visual information to our model; secondly, we fuse the representation of original RGB clips with multiple non-strictly corresponding depth clips sampled by our temporal asynchronization augmentation mechanism, which synthesizes new instances at feature-level; thirdly, a novel Depth Guided Adaptive Instance Normalization (DGAdaIN) fusion module is proposed to fuse the two-stream modalities efficiently. Additionally, to better mimic the few-shot recognition process, our model is trained in the meta-learning way. Extensive experiments on several action recognition benchmarks demonstrate the effectiveness of our model.

In this paper we illustrate how to perform both visual object tracking and semi-supervised video object segmentation, in real-time, with a single simple approach. Our method, dubbed SiamMask, improves the offline training procedure of popular fully-convolutional Siamese approaches for object tracking by augmenting their loss with a binary segmentation task. Once trained, SiamMask solely relies on a single bounding box initialisation and operates online, producing class-agnostic object segmentation masks and rotated bounding boxes at 55 frames per second. Despite its simplicity, versatility and fast speed, our strategy allows us to establish a new state-of-the-art among real-time trackers on VOT-2018, while at the same time demonstrating competitive performance and the best speed for the semi-supervised video object segmentation task on DAVIS-2016 and DAVIS-2017.

We present a conceptually simple, flexible, and general framework for few-shot learning, where a classifier must learn to recognise new classes given only few examples from each. Our method, called the Relation Network (RN), is trained end-to-end from scratch. During meta-learning, it learns to learn a deep distance metric to compare a small number of images within episodes, each of which is designed to simulate the few-shot setting. Once trained, a RN is able to classify images of new classes by computing relation scores between query images and the few examples of each new class without further updating the network. Besides providing improved performance on few-shot learning, our framework is easily extended to zero-shot learning. Extensive experiments on five benchmarks demonstrate that our simple approach provides a unified and effective approach for both of these two tasks.

Zero-shot learning (ZSL) models rely on learning a joint embedding space where both textual/semantic description of object classes and visual representation of object images can be projected to for nearest neighbour search. Despite the success of deep neural networks that learn an end-to-end model between text and images in other vision problems such as image captioning, very few deep ZSL model exists and they show little advantage over ZSL models that utilise deep feature representations but do not learn an end-to-end embedding. In this paper we argue that the key to make deep ZSL models succeed is to choose the right embedding space. Instead of embedding into a semantic space or an intermediate space, we propose to use the visual space as the embedding space. This is because that in this space, the subsequent nearest neighbour search would suffer much less from the hubness problem and thus become more effective. This model design also provides a natural mechanism for multiple semantic modalities (e.g.,~attributes and sentence descriptions) to be fused and optimised jointly in an end-to-end manner. Extensive experiments on four benchmarks show that our model significantly outperforms the existing models.

Most existing person re-identification (re-id) methods focus on learning the optimal distance metrics across camera views. Typically a person's appearance is represented using features of thousands of dimensions, whilst only hundreds of training samples are available due to the difficulties in collecting matched training images. With the number of training samples much smaller than the feature dimension, the existing methods thus face the classic small sample size (SSS) problem and have to resort to dimensionality reduction techniques and/or matrix regularisation, which lead to loss of discriminative power. In this work, we propose to overcome the SSS problem in re-id distance metric learning by matching people in a discriminative null space of the training data. In this null space, images of the same person are collapsed into a single point thus minimising the within-class scatter to the extreme and maximising the relative between-class separation simultaneously. Importantly, it has a fixed dimension, a closed-form solution and is very efficient to compute. Extensive experiments carried out on five person re-identification benchmarks including VIPeR, PRID2011, CUHK01, CUHK03 and Market1501 show that such a simple approach beats the state-of-the-art alternatives, often by a big margin.