Unsupervised traffic anomaly detection aims to detect abnormal frames and objects in dashcam videos given only normal videos during training. Most prior works focus on reconstruction or prediction pipelines, while recent graph-based extensions still rely on graph feature learning or limited statistical interactions. Consequently, reasoning over structured relations and abnormal interactions remains largely unexplored. This paper introduces a structured graph reasoning method that bridges normality learning, anomaly detection, and structured reasoning in a unified framework. Unlike prior works, our approach interprets the graph as a semantic abstraction tailored for comprehensive traffic scene understanding. It includes three innovative components: (1) A domain-specific graph representation that models semantic relationships for road scenarios, providing a holistic description of traffic activities. (2) A graph reasoning detector that detects traffic anomalies by reasoning over structured traffic scene graphs. It integrates node anomaly, i.e., unusual object movements, and edge anomaly, i.e., abnormal object interactions, within a unified graph optimization approach, enabling the effective detection of complex, interaction-driven anomalies. (3) An Anomaly Transformer that predicts future traffic scene graphs by attending node and edge queries to object locations and relation histories, thereby effectively learning individual and interaction-level normal traffic behaviors. Experiments on two public datasets, i.e., DoTA and MM-AU, demonstrate the superior performance of our method and its effectiveness in detecting complex abnormal traffic events.

Developing generalizable robotic policies that balance inference efficiency, manipulation accuracy, and robustness remains a formidable challenge. Existing Vision-Language-Action models demand prohibitive data scales, while keyframe-based approaches struggle to reconcile the expressivity of generative models with the latency of iterative sampling. To address this trilemma, we present Flow2Act, a unified framework that integrates agglomerative perception with a deterministic one-step generative policy. Unlike prior methods relying on separate semantic encoders or iterative diffusion processes, our approach introduces three key innovations. First, we employ an agglomerative multi-teacher visual backbone that distills complementary strengths from diverse foundation models, capturing semantics, spatial structure, and segmentation to yield robust representations without task-specific pretraining. Second, we propose a conditional MeanFlow policy that parameterizes the interval-averaged velocity field. This formulation enables genuine single-step action generation, eliminating the discretization errors and computational overhead inherent in ODE-based flow matching. Third, we devise a curriculum region-aware mechanism via a Spatial-Grounded State Space Duality architecture, which progressively shifts attention from global flow stability to fine-grained contact precision. We evaluate Flow2Act on challenging simulation benchmarks and real-world robotic tasks, demonstrating significant gains in policy performance, robustness to environmental perturbations, and cross-task real-world applicability. Videos, code, and more details are available at project page.

Video contrastive learning (V-CL) has emerged as a popular framework for unsupervised video representation learning, demonstrating strong results in tasks such as action classification and detection. Yet, to harness these benefits, it is critical for the learned representations to fully capture both static and dynamic semantics. However, our experiments show that existing V-CL methods fail to effectively learn either type of feature. Through a rigorous theoretical analysis based on the Structural Causal Model and gradient update, we find that in a given dataset, certain static semantics consistently co-occur with specific dynamic semantics. This phenomenon creates spurious correlations between static and dynamic semantics in the dataset. However, existing V-CL methods do not differentiate static and dynamic similarities when computing sample similarity. As a result, learning only one type of semantics is sufficient for the model to minimize the contrastive loss. Ultimately, this causes the V-CL pre-training process to prioritize learning the easier-to-learn semantics. To address this limitation, we propose Bi-level Optimization with Decoupling for Video Contrastive Learning. (BOD-VCL). In BOD-VCL, we model videos as linear dynamical systems based on Koopman theory. In this system, all frame-to-frame transitions are represented by a linear Koopman operator. By performing eigen-decomposition on this operator, we can separate time-variant and time-invariant components of semantics, which allows us to explicitly separate the static and dynamic semantics in the video. By modeling static and dynamic similarity separately, both types of semantics can be fully exploited during the V-CL training process. BOD-VCL can be seamlessly integrated into existing V-CL frameworks, and experimental results highlight the significant improvements achieved by our method. The source code is released at https://github.com/ZeenSong/Video_contrastive .

Dense video captioning aims to identify and generate descriptions for all events in untrimmed videos. Recently, several methods have utilized ground-truth captions to construct the retrieval corpus, enhancing video features with semantic cues to improve the quality of the generated captions. However, employing full sentences as retrieval targets may lead to retrieval results that are semantically redundant and inconsistent with the video content. Such redundant information may interfere with the visual modality during cross-modal alignment. To address this issue, we propose the Action Retrieval Dense Video Captioning (ARDVC) approach. Instead of retrieving full sentences, we extract the action-relevant descriptions from the input sentences and construct an Action-Centric Text Corpus for retrieval. Additionally, to mitigate the impact of redundant visual information, we propose the Action-Event Queries Cross Decoder, which refines frame-level features by integrating action-discriminative cues, thereby guiding the model to focus on action-relevant visual content. The integration of two components enables the model to better attend to discriminative action content, thereby generating more accurate video captions. Experiments on the ActivityNet Captions and YouCook2 datasets show that ARDVC achieves state-of-the-art performance on dense video captioning. Further qualitative analysis demonstrates that our approach efficiently eliminates semantic redundancy in both visual and textual modalities.

Understanding social interactions from visual cues is a fundamental challenge for a socially competent AI. While powerful pre-trained vision-language models (VLMs) have shown remarkable general capabilities, they surprisingly struggle to unify and learn multiple social perception tasks simultaneously, often exhibiting negative transfer. We identify that this negative transfer stems from a critical issue we term "social degradation," whereby the general visual-linguistic pre-training process of VLMs impairs the visual encoder's ability to represent nuanced social information. We investigate this behavior further under two lenses: decodability through linear representation probing and compatibility through gradient conflict analysis, revealing that both play a role in the degradation, especially the former, which is significantly compromised in the VLM pre-training process. To address these issues, we propose SocialFusion, a unified framework that learns a minimal connection between a frozen visual encoder and a language model. Compared with existing VLMs, it exhibits positive transfer across all five social tasks, leveraging synergies between them to enhance overall performance and achieves comparable performance to task-specific state-of-the-art models on various benchmarks. Our findings suggest that current VLM pre-training strategies may be detrimental to acquiring general social competence and highlight the need for more socially-aware training paradigms.

Predicting human future trajectories is crucial for various intelligent systems and applications. Previous approaches typically adopt a direct prediction strategy, which decodes trajectory features directly into future coordinates. However, they overlook different hierarchical high-order velocities, which have stronger representational abilities in dynamics. In this paper, we introduce HDDNet, a novel trajectory prediction framework that follows dynamical principles and employs a hierarchical dynamics decomposition strategy. Specifically, HDDNet models future trajectories by progressively transferring trajectory coordinates into velocity, acceleration, and jerk, up to the highest-order velocity, which sequentially represent a broader receptive field and a more compact representation of motion dynamics. Furthermore, we design a hierarchical dynamics decomposition decoder with a corresponding dynamics loss, which predicts future trajectories by sequentially refining human motions from the highest-order velocity down to the final coordinates. Compared to the traditional direct prediction strategy, our approach makes better use of dynamic information at different levels. Extensive experiments and ablation studies on the ETH-UCY, SDD and GigaTraj datasets demonstrate that our method outperforms existing state-of-the-art approaches.

Multi-view 3D detection with bird's eye view (BEV) is crucial for autonomous driving and robotics, but its robustness in real-world is limited as it struggles to predict accurate depth values. A mainstream solution, cross-modal distillation, transfers depth information from LiDAR to camera models but also unintentionally transfers depth-irrelevant information (e.g. LiDAR density). To mitigate this issue, we propose RayD3D, which transfers crucial depth knowledge along the ray: a line projecting from the camera to true location of an object. It is based on the fundamental imaging principle that predicted location of this object can only vary along this ray, which is finally determined by predicted depth value. Therefore, distilling along the ray enables more effective depth information transfer. More specifically, we design two ray-based distillation modules. Ray-based Contrastive Distillation (RCD) incorporates contrastive learning into distillation by sampling along the ray to learn how LiDAR accurately locates objects. Ray-based Weighted Distillation (RWD) adaptively adjusts distillation weight based on the ray to minimize the interference of depth-irrelevant information in LiDAR. For validation, we widely apply RayD3D into three representative types of BEV-based models, including BEVDet, BEVDepth4D, and BEVFormer. Our method is trained on clean NuScenes, and tested on both clean NuScenes and RoboBEV with a variety types of data corruptions. Our method significantly improves the robustness of all the three base models in all scenarios without increasing inference costs, and achieves the best when compared to recently released multi-view and distillation models.

Embodied dialogue localization aims to determine a target location on a given 2D map based on dialogues. This task is critical for various real-world applications, such as emergency search and rescue, where precise localization is essential. The model must focus on fine features of the 2D map, guided by multi-round natural language dialogues. While previous research has yielded satisfactory results in a rough range, practical applications demand more precise localization to facilitate navigation or object manipulation by embodied agents. To address this challenge, we introduce an Iterative Refinement and Adaptive Multi-granularity Perception network, namely RAMP, which aims to iteratively refine the target location while enhancing the interaction of features at different granularities. Experimental results on the WAY dataset show that our method outperforms the state-of-the-art methods in both single-shot (+29.34%) and multi-shot (+36.85%) settings. These results highlight the superior performance of RAMP and its significant advancement over existing models.

Hair editing is a long-standing problem in computer vision that demands both fine-grained local control and intuitive user interactions across diverse modalities. Despite the remarkable progress of GANs and diffusion models, existing methods still lack a unified framework that simultaneously supports arbitrary interaction modes (e.g., text, sketch, mask, and reference image) while ensuring precise editing and faithful preservation of irrelevant attributes. In this work, we introduce a novel paradigm that reformulates hair editing as proxy-based hair transfer. Specifically, we leverage the dense and semantically disentangled latent space of StyleGAN for precise manipulation and exploit its feature space for disentangled attribute preservation, thereby decoupling the objectives of editing and preservation. Our framework unifies different modalities by converting editing conditions into distinct transfer proxies, whose features are seamlessly blended to achieve global or local edits. Beyond 2D, we extend our paradigm to 3D-aware settings by incorporating EG3D and PanoHead, where we propose a multi-view boosted hair feature localization strategy together with 3D-tailored proxy generation methods that exploit the inherent properties of 3D-aware generative models. Extensive experiments demonstrate that our method consistently outperforms prior approaches in editing effects, attribute preservation, visual naturalness, and multi-view consistency, while offering unprecedented support for multimodal and mixed-modal interactions.

Recent feed-forward models have significantly advanced geometry perception for inferring dense 3D structure from sensor observations. However, its essential capabilities remain fragmented across multiple incompatible paradigms, including online perception, offline reconstruction, multi-modal integration, long-horizon scalability, and metric-scale estimation. We present UniT, a unified model built upon a novel Group Autoregressive Transformer, which reformulates these seemingly disparate capabilities within a single framework. The key idea is to treat groups of sensor observations as the basic autoregressive units and predict the corresponding point maps in an anchor-free and scale-adaptive manner. More specifically, diverse view configurations in both online and offline settings are naturally unified within a single group autoregression process. By varying the group size, online mode operates over multiple autoregressive steps with single-frame groups, whereas offline mode aggregates a multi-frame group in a single forward pass. Meanwhile, a queue-style KV caching mechanism ensures bounded autoregressive memory over long horizons. This is enabled by reducing long-range dependencies on early frames through anchor-free relational modeling, thereby allowing outdated memory to be discarded on the fly. To improve metric-scale generalization across scenes, a scale-adaptive geometry loss is further introduced within this framework. It couples relative geometric constraints with a partial absolute scale term, implicitly regularizing global scale and inducing a progressive transition from scale-invariant geometry to metric-scale solutions. Together with a dedicated modal attention module for integrating auxiliary modalities, UniT achieves state-of-the-art performance in unified geometry perception, as validated on ten benchmarks spanning seven representative tasks.

Recent image-goal navigation (ImageNav) methods learn a perception-action policy by separately capturing semantic features of the goal and egocentric images, then passing them to a policy network. However, challenges remain: (1) Semantic features often fail to provide accurate directional information, leading to superfluous actions, and (2) performance drops significantly when viewpoint inconsistencies arise between training and application. To address these challenges, we propose RSRNav, a simple yet effective method that reasons spatial relationships between the goal and current observations as navigation guidance. Specifically, we model the spatial relationship by constructing correlations between the goal and current observations, which are then passed to the policy network for action prediction. These correlations are progressively refined using fine-grained cross-correlation and direction-aware correlation for more precise navigation. Extensive evaluation of RSRNav on three benchmark datasets demonstrates superior navigation performance, particularly in the "user-matched goal" setting, highlighting its potential for real-world applications.

Monocular 3D object detection (M3OD) is intrinsically ill-posed, hence training a high-performance deep learning based M3OD model requires a humongous amount of labeled data with complicated visual variation from diverse scenes, variety of objects and camera poses. However, we observe that, due to strong human bias, the three independent entities, i.e., object, scene, and camera pose, are always tightly entangled when an image is captured to construct training data. More specifically, specific 3D objects are always captured in particular scenes with fixed camera poses, and hence lacks necessary diversity. Such tight entanglement induces the challenging issues of insufficient utilization and overfitting to uniform training data. To mitigate this, we propose an online object-scene-camera decomposition and recomposition data manipulation scheme to more efficiently exploit the training data. We first fully decompose training images into textured 3D object point models and background scenes in an efficient computation and storage manner. We then continuously recompose new training images in each epoch by inserting the 3D objects into the freespace of the background scenes, and rendering them with perturbed camera poses from textured 3D point representation. In this way, the refreshed training data in all epochs can cover the full spectrum of independent object, scene, and camera pose combinations. This scheme can serve as a plug-and-play component to boost M3OD models, working flexibly with both fully and sparsely supervised settings. In the fully-supervised setting, all objects are annotated. In the sparsely-supervised setting, objects closest to the ego-camera for all instances are sparsely annotated. We then can flexibly increase the annotated objects to control annotation cost. For validation, our method is widely applied to five representative M3OD models and evaluated on both the KITTI and the more complicated Waymo datasets. In the fully-supervised setting, our method significantly improves the performance of the base models by 26 ∼ 48

Pedestrian trajectory prediction is crucial for ensuring safe decision-making in intelligent robotic systems. While this task demands real-time performance, previous works have primarily focused on improving prediction accuracy, often neglecting efficiency. Dense predictions with time-consuming post-clustering steps and global interactions with quadratic computational complexity result in a trade-off between accuracy and speed. In this paper, we propose a novel Sparse Trajectory Prediction (STP) model that aims to achieve both high accuracy and real-time speed by following an efficient principle: leveraging sparse structures to achieve global effects. STP instantiates this principle within a transformer-style encoder-decoder framework. In the encoder, STP introduces irregular interaction, which builds sparse interactions with dynamic interactive positions, reducing computational complexity to linearithmic/linear while maintaining global interaction. In the decoder, STP applies an early-sparsity strategy to generate sparse motion modes that represent global motion behaviors. These modes are shared across all predictions, eliminating redundant computations. By harnessing the expressive power of transformers, STP maps these sparse motion modes into multimodal future trajectories, significantly improving prediction speed while ensuring accuracy. Experimental results on four commonly used datasets demonstrate that STP maximizes both accuracy and prediction speed, achieving state-of-the-art performance and significantly improving prediction speed by about $100 \times$100× - $150 \times$150× to satisfy the real-time demand.

Weakly-supervised object detection (WSOD) learns detectors with only image-level classification annotations. Without precise instance-level labels, most previous WSOD methods in remote sensing images (RSIs) select the highest-scoring proposals as the final detection results, which are confronted by two major challenges: (1) instances with small scale or rare poses are easily neglected; (2) optimizing network by the top-scoring region inevitably overlooks many valuable candidate proposals. To mitigate the above-mentioned challenges, we propose a data-driven bidirectional spatial-adaptive network (BSANet). It contains a forward-reverse spatial dropout (FRSD) module to reduce instance ambiguity induced from extreme scales and poses, as well as crowded scene, and to better excavate the entire instances. From attention learning perspective, the proposed FRSD is conceptually similar to a data-driven hard attention mechanism, which adaptively samples and reconstructs the spatially related regions for mining more latent feature responses. Meanwhile, our FRSD effectively alleviates the inherent problem that non-parametric hard attention learning fashion cannot adapt to different datasets. In addition, we build a soft attention branch to simultaneously model soft pixel-level and hard region-level attention information for exploring the complementary benefit between soft and hard attention learning. We evaluate our BSANet on the challenging NWPU VHR-10.v2 and DIOR datasets. Experimental results demonstrate that our method sets a new state-of-the-art.

Despite Multimodal Large Language Models (MLLMs) having shown impressive capabilities, they may suffer from hallucinations. Empirically, we find that MLLMs attend disproportionately to task-irrelevant background regions compared with text-only LLMs, implying spurious background-answer correlations. We claim and analyze that (i) outcome-based rewards can be an important factor leading to spurious correlations, and (ii) spurious correlations can be an important factor leading to hallucinations. Based on these results, we propose Causal-Oriented Policy Optimization (COPO) to mitigate these spurious correlations, thus addressing the issue of hallucinations. It imposes token-level sufficiency and necessity constraints to measure each inference token's causal contribution, thus ensuring correct and evidence-grounded output. Specifically, we first evaluate each token's causal contribution via a newly proposed causal completeness reward. This reward is then used to construct a causally informed advantage function within the GRPO optimization framework, encouraging the model to focus on tokens that are causally sufficient and necessary for accurate generation. Experimental results across various benchmarks demonstrate the advantages of COPO.

Multi-camera 3D object detection (MC3D) has attracted increasing attention with the growing deployment of multi-sensor physical agents, such as robots and autonomous vehicles. However, MC3D models still struggle to generalize to unseen platforms with new multi-camera configurations. Current solutions simply employ a meta-camera for unified representation but lack comprehensive consideration. In this paper, we revisit this issue and identify that the devil lies in spatial prior discrepancies across source and target configurations, including different intrinsics, extrinsics, and array layouts. To address this, we propose CoIn3D, a generalizable MC3D framework that enables strong transferability from source configurations to unseen target ones. CoIn3D explicitly incorporates all identified spatial priors into both feature embedding and image observation through spatial-aware feature modulation (SFM) and camera-aware data augmentation (CDA), respectively. SFM enriches feature space by integrating four spatial representations, such as focal length, ground depth, ground gradient, and Plücker coordinate. CDA improves observation diversity under various configurations via a training-free dynamic novel-view image synthesis scheme. Extensive experiments demonstrate that CoIn3D achieves strong cross-configuration performance on landmark datasets such as NuScenes, Waymo, and Lyft, under three dominant MC3D paradigms represented by BEVDepth, BEVFormer, and PETR.

Action-conditioned world models are useful for embodied agents only when their predicted futures remain controllable by actions and stable under long-horizon rollout. We present SAMPO++, an embodied world model that treats action-conditioned prediction as a scale-decoupled controlled dynamical system rather than as conventional action-conditioned video generation. SAMPO++ couples temporal autoregression with scale-wise flow matching in a continuous latent pyramid. A multi-scale temporal planner summarizes the latent history with scale-matched recurrent states, allowing dynamics at different resolutions to be conditioned by appropriate temporal contexts. An Action-Controlled Velocity Field (ACVF) separates action free passive drift from action-induced residual dynamics, making the action an explicit bounded control input to the flow velocity instead of a passive conditioning token. To improve cross scale consistency and closed-loop stability, SAMPO++ further uses Pyramid-Consistent RoPE (PC-RoPE) and rollout-aware training, aligning scale-wise coordinates and exposing the model to its own autoregressive prediction distribution during training. Beyond standard perceptual metrics such as FVD, PSNR, SSIM, and LPIPS, we evaluate world-model-native properties including action alignment, counterfactual accuracy, no-op residual, and rollout drift. We further study SAMPO++ as a learned simulator for visual planning and policy-level evaluation, where policies or action chunks can be rolled out and compared before execution. Experiments across action-conditioned robotic manipulation, visual planning, and model-based reinforcement learning, together with action-free driving video prediction, show that SAMPO++ improves visual prediction quality while providing stronger action alignment, counterfactual accuracy, no-op residual suppression, and long-horizon rollout consistency over strong discrete and continuous baselines. Videos and additional details are available on the project page.

Pedestrian trajectory prediction aims to forecast potential movements of pedestrians given their historical observations, which plays an important role in ensuring safety of autonomous driving and robot navigation in social scenes. Plenty of works have achieved great advances when sufficient historical information is available. However, in many real-world cases in which observations are incomplete or limited due to tracking losses or sudden emergence of agents. Although recent works attempt to address this issue by designing momentary observation feature extractors or performing bi-directional prediction using diffusion to assist momentary feature learning, they are sub-optimal in prediction accuracy or computationally inefficient. In this paper, we propose a posterior augmented conditional variational autoencoder to perform momentary trajectory prediction. Specifically, we introduce unobserved historical features in the posterior encoder in CVAE, which enforces the model to utilize useful unobserved features to facilitate momentary feature learning and prediction. Moreover, reverse-enhanced multi-head attention is leveraged to make the model enhance the focus of the model on different but supplementary parts of unobserved features concerning momentary observation. This enables the model to learn more comprehensive representations, ultimately improving its ability to assist in accurate momentary predictions. Extensive experiments demonstrate that our model achieves state-of-the-art performance on ETH-UCY and Stanford Drone Datasets and our proposed module does help pedestrian trajectory prediction with momentary observations.