ICCV 2025 Papers

In this work, we propose a framework that creates a lively virtual dynamic scene with contextual motions of multiple humans. Generating multi-human contextual motion requires holistic reasoning over dynamic relationships among human-human and human-scene interactions. We adapt the power of a large language model (LLM) to digest the contextual complexity within textual input and convert the task into tangible subproblems such that we can generate multi-agent behavior beyond the scale that was not considered before. Specifically, our event generator formulates the temporal progression of a dynamic scene into a sequence of small events. Each event calls for a well-defined motion involving relevant characters and objects. Next, we synthesize the motions of characters at positions sampled based on spatial guidance. We employ a high-level module to deliver scalable yet comprehensive context, translating events into relative descriptions that enable the retrieval of precise coordinates. As the first to address this problem at scale and with diversity, we offer a benchmark to assess diverse aspects of contextual reasoning. Benchmark results and user studies show that our framework effectively captures scene context with high scalability.

Modern over-parameterized deep models are highly data-dependent, with large scale general-purpose and domain-specific datasets serving as the bedrock for rapid advancements. However, many datasets are proprietary or contain sensitive information, making unrestricted model training problematic. In the open world where data thefts cannot be fully prevented, Dataset Ownership Verification (DOV) has emerged as a promising method to protect copyright by detecting unauthorized model training and tracing illicit activities. Due to its diversity and superior stealth, evading DOV is considered extremely challenging. However, this paper identifies that previous studies have relied on oversimplistic evasion attacks for evaluation, leading to a false sense of security. We introduce a unified evasion framework, in which a teacher model first learns from the copyright dataset and then transfers task-relevant yet identifier-independent domain knowledge to a surrogate student using an out-of-distribution (OOD) dataset as the intermediary. Leveraging Vision-Language Models and Large Language Models, we curate the most informative and reliable subsets from the OOD gallery set as the final transfer set, and propose selectively transferring task-oriented knowledge to achieve a better trade-off between generalization and evasion effectiveness. Experiments across diverse datasets covering eleven DOV methods demonstrate our approach simultaneously eliminates all copyright identifiers and significantly outperforms nine state-of-the-art evasion attacks in both generalization and effectiveness, with moderate computational overhead. As a proof of concept, we reveal key vulnerabilities in current DOV methods, highlighting the need for long-term development to enhance practicality.

3D Gaussian Splatting (3D-GS) has demonstrated exceptional capabilities in synthesizing novel views of 3D scenes. However, its training is heavily reliant on high-quality images and precise camera poses. Meeting these criteria can be challenging in non-ideal real-world conditions, where motion-blurred images frequently occur due to high-speed camera movements.To address these challenges, we introduce Event Stream Assisted Gaussian Splatting (EvaGaussians), a novel approach that harnesses event streams captured by event cameras to facilitate the learning of high-quality 3D-GS from motion-blurred images. Capitalizing on the high temporal resolution and dynamic range offered by event streams, we seamlessly integrate them into the initialization and optimization of 3D-GS, thereby enhancing the acquisition of high-fidelity novel views with intricate texture details. We also contribute two novel datasets comprising RGB frames, event streams, and corresponding camera parameters, featuring a wide variety of scenes and various camera motions. The comparison results reveal that our approach not only excels in generating high-fidelity novel views, but also offers improved training and inference efficiency. Video and code are available at the project page.

Event-based cameras (EBCs) have emerged as a bio-inspired alternative to traditional cameras, offering advantages in power efficiency, temporal resolution, and high dynamic range. However, the development of image analysis methods for EBCs is challenging due to the sparse and asynchronous nature of the data. This work addresses the problem of object detection for EBC cameras. The current approaches to EBC object detection focus on constructing complex data representations and rely on specialized architectures. We introduce I2EvDet (Image-to-Event Detection), a novel adaptation framework that bridges mainstream object detection with temporal event data processing. First, we demonstrate that a Real-Time DEtection TRansformer, or RT-DETR, a state-of-the-art natural image detector, trained on a simple image-like representation of the EBC data achieves performance comparable to specialized EBC methods. Next, as part of our framework, we develop an efficient adaptation technique that transforms image-based detectors into event-based detection models by modifying their frozen latent representation space through minimal architectural additions. The resulting EvRT-DETR model reaches state-of-the-art performance on the standard benchmark datasets Gen1 (mAP +2.3) and 1Mpx/Gen4 (mAP +1.4). These results demonstrate a fundamentally new approach to EBC object detection through principled adaptation of mainstream architectures, offering an efficient alternative with potential applications to other temporal visual domains.

Estimating 2D camera motion is a fundamental computer vision task that models the projection of 3D camera movements onto the 2D image plane. Current methods rely on either homography-based approaches, limited to planar scenes, or meshflow techniques that use grid-based local homographies but struggle with complex non-linear transformations. We introduce CamFlow, a novel framework that represents camera motion using hybrid motion bases: physical bases derived from camera geometry and stochastic bases for complex scenarios. Our approach includes a hybrid probabilistic loss function based on the Laplace distribution that enhances training robustness. For evaluation, we create a new benchmark by masking dynamic objects in existing optical flow datasets to isolate pure camera motion. Experiments show CamFlow outperforms state-of-the-art methods across diverse scenarios, demonstrating superior robustness and generalization in zero-shot settings. Code and datasets are available at our project page: https://lhaippp.github.io/CamFlow/.

Understanding user actions and their possible mistakes is essential for successful operation of task assistants. In this paper, we develop a unified framework for joint temporal action segmentation and error recognition (recognizing when and which type of error happens) in procedural task videos. We propose a Generalized Task Graph (GTG) whose nodes encode correct steps and background (task-irrelevant actions). We then develop a GTG-Video Alignment algorithm (GTG2Vid) to jointly segment videos into actions and detect frames containing errors. Given that it is infeasible to gather many videos and their annotations for different types of errors, we study a framework that only requires normal (error-free) videos during training. More specifically, we leverage large language models (LLMs) to obtain error descriptions and subsequently use video-language models (VLMs) to generate visually-aligned textual features, which we use for error recognition. We then propose an Error Recognition Module (ERM) to recognize the error frames predicted by GTG2Vid using the generated error features. By extensive experiments on two egocentric datasets of EgoPER and CaptainCook4D, we show that our framework outperforms other baselines on action segmentation, error detection and recognition.

Concept erasure techniques have recently gained significant attention for their potential to remove unwanted concepts from text-to-image models. While these methods often demonstrate promising results in controlled settings, their robustness in real-world applications and suitability for deployment remain uncertain. In this work, we (1) identify a critical gap in evaluating sanitized models, particularly in assessing their performance across diverse concept dimensions, and (2) systematically analyze the failure modes of text-to-image models post-erasure. We focus on the unintended consequences of concept removal on non-target concepts across different levels of interconnected relationships including visually similar, binomial, and semantically related concepts. To address this, we introduce EraseBench, a comprehensive benchmark for evaluating post-erasure performance. EraseBench includes over 100 curated concepts, targeted evaluation prompts, and a robust set of metrics to assess both effectiveness and side effects of erasure. Our findings reveal a phenomenon of concept entanglement, where erasure leads to unintended suppression of non-target concepts, causing spillover degradation that manifests as distortions and a decline in generation quality.

Pre-trained vision foundation models have transformed many computer vision tasks. Despite their strong ability to learn discriminative and generalizable features crucial for out-of-distribution (OOD) detection, their impact on this task remains underexplored. Motivated by this gap, we systematically investigate representative vision foundation models for OOD detection. Our findings reveal that a pre-trained DINOv2 model, even without fine-tuning on in-domain (ID) data, naturally provides a highly discriminative feature space for OOD detection, achieving performance comparable to existing state-of-the-art methods without requiring complex designs. Beyond this, we explore how fine-tuning foundation models on in-domain (ID) data can enhance OOD detection. However, we observe that the performance of vision foundation models remains unsatisfactory in scenarios with a large semantic space. This is due to the increased complexity of decision boundaries as the number of categories grows, which complicates the optimization process. To mitigate this, we propose the Mixture of Feature Experts (MoFE) module, which partitions features into subspaces, effectively capturing complex data distributions and refining decision boundaries. Further, we introduce a Dynamic-b Mixup strategy, which samples interpolation weights from a dynamic beta distribution. This adapts to varying levels of learning difficulty across categories, improving feature learning for more challenging categories. Extensive experiments demonstrate the effectiveness of our approach, significantly outperforming baseline methods.

Learning self-supervised representations that are invariant and equivariant to transformations is crucial for advancing beyond traditional visual classification tasks. However, many methods rely on predictor architectures to encode equivariance, despite evidence that architectural choices, such as capsule networks, inherently excel at learning interpretable pose-aware representations. To explore this, we introduce EquiCaps (Equivariant Capsule Network), a capsule-based approach to pose-aware self-supervision that eliminates the need for a specialised predictor for enforcing equivariance. Instead, we leverage the intrinsic pose-awareness capabilities of capsules to improve performance in pose estimation tasks. To further challenge our assumptions, we increase task complexity via multi-geometric transformations to enable a more thorough evaluation of invariance and equivariance by introducing 3DIEBench-T, an extension of a 3D object-rendering benchmark dataset. Empirical results demonstrate that EquiCaps outperforms prior state-of-the-art equivariant methods on geometric tasks, including rotation and translation, achieving a supervised-level R^2 of 0.78 on the 3DIEBench rotation prediction benchmark and improving upon SIE and CapsIE by 0.05 and 0.04 R^2, respectively. Moreover, in contrast to non-capsule-based equivariant approaches, EquiCaps maintains robust equivariant performance under combined geometric transformations, underscoring its generalisation capabilities and the promise of predictor-free capsule architectures. Code, dataset, and weights are released at http://github.com/AberdeenML/EquiCaps.

Diffusion models have demonstrated exceptional visual quality in video generation, making them promising for autonomous driving world modeling. However, existing video diffusion-based world models struggle with flexible-length, long-horizon predictions and integrating trajectory planning. This is because conventional video diffusion models rely on global joint distribution modeling of fixed-length frame sequences rather than sequentially constructing localized distributions at each timestep. In this work, we propose Epona, an autoregressive diffusion world model that enables localized spatiotemporal distribution modeling through two key innovations: 1) Decoupled spatiotemporal factorization that separates temporal dynamics modeling from fine-grained future world generation, and 2) Modular trajectory and video prediction that seamlessly integrate motion planning with visual modeling in an end-to-end framework. Our architecture enables high-resolution, long-duration generation while introducing a novel chain-of-forward training strategy to address error accumulation in autoregressive loops. Experimental results demonstrate state-of-the-art performance with 7.4% FVD improvement and minutes-longer prediction duration compared to prior works. The learned world model further serves as a real-time motion planner, outperforming strong end-to-end planners on NAVSIM benchmarks.

Disparity estimation is an essential step in processing and analyzing Light Field (LF) images. Recent methods construct the cost volume to exploit the correspondence of the LFs over the preset maximum disparity, limiting them to process the large parallax scenes. Different from constructing cost volume, the self-attention mechanism calculates the parallax attention between epipolar lines to find the matching points. However, for LFs that have different views, the related disparity scales are different in parallax attention since the baselines with the central view are different. Moreover, if the matching information is occluded in one view, the disparity information can be explored through other views. Therefore, mapping these attentions to the same scale and selecting effective matching information are key points for disparity estimation from parallax attention. In this paper, we explore parallax attention for LF and design an unsupervised method, named Epipolar Consistent Attention Aggregation Network (ECAAN). We first introduce an epipolar consistent scale unification block by considering the consistency relationships to standardize disparity scales of the parallax attention maps. Based on the intra-properties and inter-relationships of parallax attention, we further propose a consistent occlusion-free aggregation block to integrate the information from the occlusion-free areas. Besides, we design an improved photometric loss to constrain the model. ECAAN achieves state-of-the-art performance in LF depth estimation. Notably, ECAAN attains a mean square error (MSE) of 0.2 on large-disparity LF datasets, achieving a 68% error reduction compared to the second-best method.

This paper introduces EA6D, a novel diffusion-based framework for 6D pose estimation that operates effectively in any environment. Traditional pose estimation methods struggle with the variability and complexity of real-world scenarios, often leading to overfitting on controlled datasets and poor generalization to new scenes. To address these challenges, we propose a generative pose estimation paradigm that generates environment-independent object representations for pose estimation, which are robust to environmental variations such as illumination, occlusion, and background clutter. Specifically, we propose the novel Environment Decoupling Diffusion Model (EDDM) which separates object representations from environmental factors while enabling efficient few-step sampling by leveraging input image priors instead of pure noise initialization. We validate our approach on four standard benchmarks and a self-made dataset DiverseScenes. The results demonstrate that EA6D, trained using only synthetic data, can outperform the state-of-the-art methods with both synthetic and realistic data. In particular, for fair comparisons with synthetic data, we can exceed the previous SOTA by 18.1% and 33.5% on Linemod and Linemod-Occluded datasets respectively. Project page: https://github.com/acmff22/EA6D

While fine-tuning diffusion models with reinforcement learning (RL) has demonstrated effectiveness in directly optimizing downstream objectives, existing RL frameworks are prone to overfitting the rewards, leading to outputs that deviate from the true data distribution and exhibit reduced diversity. To address this issue, we introduce entropy as a quantitative measure to enhance the exploratory capacity of diffusion models' denoising policies. We propose an adaptive mechanism that dynamically adjusts the application and magnitude of entropy and regularization, guided by real-time quality estimation of intermediate noised states. Theoretically, we prove the convergence of our entropy-enhanced policy optimization and establish two critical properties: 1) global entropy increases through training, ensuring robust exploration capabilities, and 2) entropy systematically decreases during the denoising process, enabling a phase transition from early-stage diversity promotion to late-stage distributional fidelity. Building on this foundation, we propose a plug-and-play RL module that adaptively controls entropy and optimizes denoising steps. Extensive evaluations demonstrate the theoretical soundness and empirical robustness of our method, achieving state-of-the-art quality-diversity trade-offs across benchmarks. Notably, our framework significantly improves the rewards and reduces denoising steps in training by up to 40%.

Unsupervised object discovery (UOD) aims to detect and segment objects in 2D images without handcrafted annotations. Recent progress in self-supervised representation learning has led to some success in UOD algorithms. However, the absence of ground truth provides existing UOD methods with two challenges: 1) determining if a discovered region is foreground or background, and 2) knowing how many objects remain undiscovered. To address these two problems, previous solutions rely on foreground priors to distinguish if the discovered region is foreground, and conduct one or fixed iterations of discovery. However, the existing foreground priors are heuristic and not always robust, and a fixed number of discoveries leads to under or over-segmentation, since the number of objects in images varies. This paper introduces UnionCut, a robust and well-grounded foreground prior based on min-cut and ensemble methods that detects the union of foreground areas of an image, allowing UOD algorithms to identify foreground objects and stop discovery once the majority of the foreground union in the image is segmented. In addition, we propose UnionSeg, a distilled transformer of UnionCut that outputs the foreground union more efficiently and accurately. Our experiments show that by combining with UnionCut or UnionSeg, previous state-of-the-art UOD methods witness an increase in the performance of single object discovery, saliency detection and self-supervised instance segmentation on various benchmarks. The code is available at https://github.com/YFaris/UnionCut.

We introduce ED-VTG, a method for fine-grained video temporal grounding utilizing multi-modal large language models. Our approach harnesses the capabilities of multimodal LLMs to jointly process text and video, in order to effectively localize natural language queries in videos through a two-stage process. Rather than being directly grounded, language queries are initially transformed into enriched sentences that incorporate missing details and cues to aid in grounding. In the second stage, these enriched queries are grounded, using a lightweight decoder, which specializes at predicting accurate boundaries conditioned on contextualized representations of the enriched queries. To mitigate noise and reduce the impact of hallucinations, our model is trained with a multiple-instance-learning objective that dynamically selects the optimal version of the query for each training sample. We demonstrate state-of-the-art results across various benchmarks in temporal video grounding and paragraph grounding settings. Experiments reveal that our method significantly outperforms all previously proposed LLM-based temporal grounding approaches and is either superior or comparable to specialized models, while maintaining a clear advantage against them in zero-shot evaluation scenarios.

Existing deep learning-based models for remote sensing pansharpening exhibit exceptional performance on training datasets. However, due to sensor-specific characteristics and varying imaging conditions, these models suffer from substantial performance degradation when applied to unseen satellite data, lacking generalizability and thus limiting their applicability. We argue that the performance drops stem primarily from distributional discrepancies from different sources and the key to addressing this challenge lies in bridging the gap between training and testing distributions. To validate the idea and further achieve a "train once, deploy forever" capability, this paper introduces a novel and intuitive approach to enpower any pansharpening models with generalizability by employing a unified distribution strategy (UniPAN). Specifically, we construct a distribution transformation function that normalizes the pixels sampled from different sources to conform to an identical distribution. The deep models are trained on the transformed domain, and during testing on new datasets, the new data are also transformed to match the training distribution. UniPAN aims to train and test the model on a unified and consistent distribution, thereby enhancing its generalizability. Extensive experiments validate the efficacy of UniPAN, demonstrating its potential to significantly enhance the performance of deep pansharpening models across diverse satellite sensors. Codes: https://github.com/yc-cui/UniPAN.

Text-guided zero-shot object counting leverages vision-language models (VLMs) to count objects of an arbitrary class given by a text prompt. Existing approaches for this challenging task only utilize local patch-level features to fuse with text feature, ignoring the important influence of the global image-level feature. In this paper, we propose a universal strategy that can exploit both local patch-level features and global image-level feature simultaneously. Specifically, to improve the localization ability of VLMs, we propose Text-guided Local Ranking. Depending on the prior knowledge that foreground patches have higher similarity with the text prompt, a new local-text rank loss is designed to increase the differences between the similarity scores of foreground and background patches which push foreground and background patches apart. To enhance the counting ability of VLMs, Number-evoked Global Attention is introduced to first align global image-level feature with multiple number-conditioned text prompts. Then, the one with the highest similarity is selected to compute cross-attention with the global image-level feature. Through extensive experiments on widely used datasets and methods, the proposed approach has demonstrated superior advancements in performance, generalization, and scalability. Furthermore, to better evaluate text-guided zero-shot object counting methods, we propose a dataset named ZSC-8K, which is larger and more challenging, to establish a new benchmark. Codes and ZSC-8K dataset will be available.

Transformers have transformed modern machine learning, driving breakthroughs in computer vision, natural language processing, and robotics. At the core of their success lies the attention mechanism, which enables the modeling of global dependencies among input tokens. However, we reveal that the attention block in transformers suffers from inherent ill-conditioning, which hampers gradient-based optimization and leads to inefficient training. To address this, we develop a theoretical framework that establishes a direct relationship between the conditioning of the attention block and that of the embedded tokenized data. Building on this insight, we introduce conditioned embedded tokens, a method that systematically modifies the embedded tokens to improve the conditioning of the attention mechanism. Our analysis demonstrates that this approach significantly mitigates ill-conditioning, leading to more stable and efficient training. We validate our methodology across various transformer architectures, achieving consistent improvements in image classification, object detection, instance segmentation, and natural language processing, highlighting its broad applicability and effectiveness.

In the field of AI security, deep neural networks (DNNs) are highly sensitive to adversarial examples (AEs), which can cause incorrect predictions with minimal input perturbations. Although AEs exhibit transferability across models, targeted attack success rates (TASRs) are low due to differences in feature dimensions and decision boundaries. To enhance targeted AE transferability, we propose a novel approach using Inverse Target Gradient Competition (ITC) and Spatial Distance Stretching (SDS) in the optimization process. Specifically, we employ a siamese-network-like framework to generate both non-targeted and targeted AEs. The ITC mechanism applies non-targeted adversarial gradients each epoch to impede the optimization of targeted perturbations, thereby improving robustness. Additionally, a top-k SDS strategy guides AEs to penetrate target class regions in the latent space while distancing from non-targeted regions, achieving optimal transferability. Compared to state-of-the-art competition-based attacks, our method significantly improves transferable TASRs by 16.1% and 21.4% on mainstream CNNs and ViTs, respectively, and demonstrates superior defense-breaking capabilities.