CVPR 2025 Papers

We introduce the first training-free approach for Video Semantic Segmentation (VSS) based on pre-trained diffusion models. A growing research direction attempts to employ diffusion models to perform downstream vision tasks by exploiting their deep understanding of image semantics. Yet, the majority of these approaches have focused on image-related tasks like semantic segmentation, with less emphasis on video tasks such as VSS. Ideally, diffusion-based image semantic segmentation approaches can be applied to videos in a frame-by-frame manner. However, we find their performance on videos to be subpar due to the absence of any modeling of temporal information inherent in the video data. To this end, we tackle this problem and introduce a framework tailored for VSS based on pre-trained image and video diffusion models. We propose building a scene context model based on the diffusion features, where the model is autoregressively updated to adapt to scene changes. This context model predicts per-frame coarse segmentation maps that are temporally consistent. To refine these maps further, we propose a correspondence-based refinement strategy that aggregates predictions temporally, resulting in more confident predictions. Finally, we introduce a masked modulation approach to upsample the coarse maps to a high-quality full resolution. Experiments show that our proposed approach significantly outperforms existing training-free image semantic segmentation approaches on various VSS benchmarks without any training or fine-tuning. Moreover, it rivals supervised VSS approaches on the VSPW dataset despite not being explicitly trained for VSS.

Multimodal large language models (MLLMs) have recently shown significant advancements in video understanding, excelling in content reasoning and instruction-following tasks. However, hallucination, where models generate inaccurate or misleading content, remains underexplored in the video domain. Building on the observation that MLLM visual encoders often fail to distinguish visually different yet semantically similar video pairs, we introduce VidHalluc, the largest benchmark designed to examine hallucinations in MLLMs for video understanding. It consists of 5,002 videos, paired to highlight cases prone to hallucinations. VidHalluc assesses hallucinations across three critical dimensions: (1) action, (2) temporal sequence, and (3) scene transition. Comprehensive testing shows that most MLLMs are vulnerable to hallucinations across these dimensions. Furthermore, we propose DINO-HEAL, a training-free method that reduces hallucinations by incorporating spatial saliency from DINOv2 to reweight visual features during inference. Our results show that DINO-HEAL consistently improves performance on VidHalluc, achieving an average improvement of 3.02% in mitigating hallucinations across all tasks. Both the VidHalluc benchmark and DINO-HEAL code are available at https://people-robots.github.io/vidhalluc.

The advancement of Multimodal Large Language Models (MLLMs) has enabled significant progress in multimodal understanding, expanding their capacity to analyze video content. However, existing evaluation benchmarks for MLLMs primarily focus on abstract video comprehension, lacking a detailed assessment of their ability to understand video compositions, the nuanced interpretation of how visual elements combine and interact within highly compiled video contexts. We introduce VidComposition, a new benchmark specifically designed to evaluate the video composition understanding capabilities of MLLMs using carefully curated compiled videos and cinematic-level annotations.VidComposition includes 982 videos with 1706 multiple-choice questions, covering various compositional aspects such as camera movement, angle, shot size, narrative structure, character actions and emotions, etc. Our comprehensive evaluation of 33 open-source and proprietary MLLMs reveals a significant performance gap between human and model capabilities. This highlights the limitations of current MLLMs in understanding complex, compiled video compositions and offers insights into areas for further improvement. Our benchmark is publicly available at https://yunlong10.github.io/VidComposition/

We present Vid2Avatar-Pro, a method to create photorealistic and animatable 3D human avatars from monocular in-the-wild videos. Building a high-quality avatar that supports animation with diverse poses from a monocular video is challenging because the observation of pose diversity and view points is inherently limited. The lack of pose variations typically leads to poor generalization to novel poses, and avatars can easily overfit to limited input view points, producing artifacts and distortions from other views. In this work, we address these limitations by leveraging a universal prior model (UPM) learned from a large corpus of multi-view clothed human performance capture data. We build our representation on top of expressive 3D Gaussians with canonical front and back maps shared across identities. Once the UPM is learned to accurately reproduce the large-scale multi-view human images, we fine-tune the model with an in-the-wild video via inverse rendering to obtain a personalized photorealistic human avatar that can be faithfully animated to novel human motions and rendered from novel views. The experiments show that our approach based on the learned universal prior sets a new state-of-the-art in monocular avatar reconstruction by substantially outperforming existing approaches relying only on heuristic regularization or a shape prior of minimally clothed bodies (e.g., SMPL) on publicly available datasets.

Recent advances in multimodal large language models (MLLMs) have expanded research in video understanding, primarily focusing on high-level tasks such as video captioning and question-answering. Meanwhile, a smaller body of work addresses dense, pixel-precise segmentation tasks, which typically involve category-guided or referral-based object segmentation. Although both research directions are essential for developing models with human-level video comprehension, they have largely evolved separately, with distinct benchmarks and architectures. This paper aims to unify these efforts by introducing ViCaS, a new dataset containing thousands of challenging videos, each annotated with detailed, human-written captions and temporally consistent, pixel-accurate masks for multiple objects with phrase grounding. Our benchmark evaluates models on both holistic/high-level understanding and language-guided, pixel-precise segmentation. We also present carefully validated evaluation measures and propose an effective model architecture that can tackle our benchmark. Project page: https://ali2500.github.io/vicas-project/

Sign language is a visual language expressed through complex movements of the upper body. The human skeleton plays a critical role in sign language recognition due to its good separation from the video background. However, mainstream skeleton-based sign language recognition models often overly focus on the natural connections between joints, treating sign language as ordinary human movements, which neglects its linguistic characteristics. We believe that just as letters form words, each sign language gloss can also be decomposed into smaller visual symbols. To fully harness the potential of skeleton data, this paper proposes a novel joint fusion strategy and a visual symbol attention model. Specifically, we first input the complete set of skeletal joints, and after dynamically exchanging joint information, we discard the parts with the weakest connections to other joints, resulting in a fused, simplified skeleton. Then, we group the joints most likely to express the same visual symbol and discuss the joint movements within each group separately. To validate the superiority of our method, we conduct extensive experiments on multiple public benchmark datasets. The results show that, without complex pre-training, we still achieve new state-of-the-art performance.

Human daily activities can be concisely narrated as sequences of routine events (e.g., turning off an alarm) in video streams, forming an event vocabulary. Motivated by this, we introduce **VLog**, a novel video understanding framework that defines video narrations as a vocabulary, going beyond the typical subword vocabularies in existing generative video-language models. Built on the lightweight language model GPT-2, **VLog** features three key innovations:1. **A Generative Retrieval Model** Marrying the language model's complex reasoning capabilities with contrastive retrieval's efficient similarity search.2. **A Hierarchical Vocabulary** Derived from large-scale video narrations using our narration pair encoding algorithm, enabling efficient indexing of specific events (e.g., cutting a tomato) by identifying broader scenarios (e.g., kitchen) with expressive postfixes (e.g., by the left hand).3. **A Vocabulary Update Strategy** Leveraging generative models to extend the vocabulary for novel events encountered during inference.To validate our approach, we introduce **VidCab-Eval**, a development set requiring concise narrations with reasoning relationships (e.g., before and after). Experiments on **EgoSchema**, **COIN**, and **HiREST** further demonstrate the effectiveness of **VLog**, highlighting its ability to generate concise, contextually accurate, and efficient narrations. This offers a novel perspective on video understanding.

We investigate complex video question answering via chain-of-evidence reasoning --- identifying sequences of temporal spans from multiple relevant parts of the video, together with visual evidence within them.Existing models struggle with multi-step reasoning as they uniformly sample a fixed number of frames, which can miss critical evidence distributed nonuniformly throughout the video. Moreover, they lack the ability to temporally localize such evidence in the broader context of the full video, which is required for answering complex questions. We propose a framework to enhance existing VideoQA datasets with evidence reasoning chains, automatically constructed by searching for optimal intervals of interest in the video with supporting evidence, that maximizes the likelihood of answering a given question.We train our model (ViTED) to generate these evidence chains directly, enabling it to both localize evidence windows as well as perform multi-step reasoning across them in long-form video content.We show the value of our evidence-distilled models on a suite of long video QA benchmarks where we outperform state-of-the-art approaches that lack evidence reasoning capabilities.

Current large multimodal models (LMMs) face significant challenges in processing and comprehending long-duration or high-resolution videos, which is mainly due to the lack of high-quality datasets. To address this issue from a data-centric perspective, we propose VISTA, a simple yet effective video spatiotemporal augmentation framework that synthesizes long-duration and high-resolution video instruction-following pairs from existing video-caption datasets. VISTA spatially and temporally combines videos to create new synthetic videos with extended durations and enhanced resolutions, and subsequently produces question-answer pairs pertaining to these newly synthesized videos. Based on this paradigm, we develop seven video augmentation methods and curate VISTA-400K, a video instruction-following dataset aimed at enhancing long-duration and high-resolution video understanding. Finetuning various video LMMs on our data resulted in an average improvement of 3.3% across four challenging benchmarks for long-video understanding. Furthermore, we introduce the first comprehensive high-resolution video understanding benchmark HRVideoBench, on which our finetuned models achieve a 6.5% performance gain. These results highlight the effectiveness of our framework.

Widely shared videos on the internet are often edited. Recently, although Video Large Language Models (Vid-LLMs) have made great progress in general video understanding tasks, their capabilities in video editing understanding (VEU) tasks remain unexplored. To address this gap, in this paper, we introduce VEU-Bench (Video Editing Understanding Benchmark), a comprehensive benchmark that categorizes video editing components across various dimensions, from intra-frame features like shot size to inter-shot attributes such as cut types and transitions. Unlike previous video editing understanding benchmarks that focus mainly on editing element classification, VEU-Bench encompasses 19 fine-grained tasks across three stages: recognition, reasoning, and judging. To enhance the annotation of VEU automatically, we built an annotation pipeline integrated with an ontology-based knowledge base. Through extensive experiments with 11 state-of-the-art Vid-LLMs, our findings reveal that current Vid-LLMs face significant challenges in VEU tasks, with some performing worse than random choice. To alleviate this issue, we develop Oscars(Named after the Academy Awards.), a VEU expert model fine-tuned on the curated VEU-Bench dataset. It outperforms existing open-source Vid-LLMs on VEU-Bench by over 28.3% in accuracy and achieves performance comparable to commercial models like GPT-4o. We also demonstrate that incorporating VEU data significantly enhances the performance of Vid-LLMs on general video understanding benchmarks, with an average improvement of 8.3% across nine reasoning tasks. The code and data will be made available.

The rapid advancement of vision-language models (VLMs) has established a new paradigm in video anomaly detection (VAD): leveraging VLMs to simultaneously detect anomalies and provide comprehendible explanations for the decisions. Existing work in this direction often assumes the complex reasoning required for VAD exceeds the capabilities of pretrained VLMs. Consequently, these approaches either incorporate specialized reasoning modules during inference or rely on instruction tuning datasets through additional training to adapt VLMs for VAD. However, such strategies often incur substantial computational costs or data annotation overhead. To address these challenges in explainable VAD, we introduce a verbalized learning framework named VERA that enables VLMs to perform VAD without model parameter modifications. Specifically, VERA automatically decomposes the complex reasoning required for VAD into reflections on simpler, more focused guiding questions capturing distinct abnormal patterns. It treats these reflective questions as learnable parameters and optimizes them through data-driven verbal interactions between learner and optimizer VLMs, using coarsely labeled training data. During inference, VERA embeds the learned questions into model prompts to guide VLMs in generating segment-level anomaly scores, which are then refined into frame-level scores via the fusion of scene and temporal contexts. Experimental results on challenging benchmarks demonstrate that the learned questions of VERA are highly adaptable, significantly improving both detection performance and explainability of VLMs for VAD.

A fundamental aspect of compositional reasoning in a video is associating people and their actions across time. Recent years have seen great progress in general-purpose vision/video models and a move towards long-video understanding. While exciting, we take a step back and ask: are today's models good at compositional reasoning on short videos? To this end, we introduce VELOCITI, a benchmark to study Video-LLMs by disentangling and assessing the comprehension of agents, actions, and their associations across multiple events. We adopt the Video-Language Entailment setup and propose StrictVLE that requires correct classification (rather than ranking) of the positive and negative caption. We evaluate several models and observe that even the best, LLaVA-OneVision (44.5%) and Gemini-1.5-Pro (49.3%), are far from human accuracy at 93.0%. Results show that action understanding lags behind agents, and negative captions created using entities appearing in the video perform worse than those obtained from pure text manipulation. We also present challenges with ClassicVLE and multiple-choice (MC) evaluation, strengthening our preference for StrictVLE. Finally, we validate that our benchmark requires visual inputs of multiple frames making it ideal to study video-language compositional reasoning.

Object permanence in humans is a fundamental cue that helps in understanding persistence of objects, even when they are fully occluded in the scene. Present day methods in object segmentation do not account for this amodal nature of the world, and only work for segmentation of visible or modal objects. Few amodal methods exist; single-image segmentation methods cannot handle high-levels of occlusions which are better inferred using temporal information, and multi-frame methods have focused solely on segmenting rigid objects. To this end, we propose to tackle video amodal segmentation by formulating it as a conditional generation task, thereby capitalizing on the foundational knowledge in video generative models. Our method is simple; we repurpose these models to condition on a sequence of modal mask frames of an object along with contextual depth maps, to learn which object boundary may be occluded and therefore, extended to hallucinate the complete extent of an object. This is followed by a content completion stage which is able to inpaint the occluded regions of an object. We benchmark our approach alongside a wide array of state-of-the-art methods on four datasets and show a dramatic improvement of upto 13% for amodal segmentation in an object's occluded region.

Existing 3D vision-language (3D-VL) benchmarks fall short in evaluating 3D-VL models, creating a "mist" that obscures rigorous insights into model capabilities and 3D-VL tasks. This mist persists due to three key limitations. First, flawed test data, like ambiguous referential text in the grounding task, can yield incorrect and unreliable test results. Second, oversimplified metrics such as simply averaging accuracy per question answering (QA) pair, cannot reveal true model capability due to their vulnerability to language variations. Third, existing benchmarks isolate the grounding and QA tasks, disregarding the underlying coherence that QA should be based on solid grounding capabilities. To unveil the "mist", we propose Beacon3D, a benchmark for 3D-VL grounding and QA tasks, delivering a perspective shift in the evaluation of 3D-VL understanding. Beacon3D features (i) high-quality test data with precise and natural language, (ii) object-centric evaluation with multiple tests per object to ensure robustness, and (iii) a novel chain-of-analysis paradigm to address language robustness and model performance coherence across grounding and QA. Our evaluation of state-of-the-art 3D-VL models on Beacon3D reveals that (i) object-centric evaluation elicits true model performance and particularly weak generalization in QA; (ii) grounding-QA coherence remains fragile in current 3D-VL models, and (iii) incorporating large language models (LLMs) to 3D-VL models, though as a prevalent practice, hinders grounding capabilities and has yet to elevate QA capabilities. We hope Beacon3D and our comprehensive analysis could benefit the 3D-VL community towards faithful developments.

Recently, text-driven video editing methods that optimize target latent representations have garnered significant attention and demonstrated promising results. However, these methods rely on self-supervised objectives to compute the gradients needed for updating latent representations, which inevitably introduces gradient noise, compromising content generation quality. Additionally, it is challenging to determine the optimal stopping point for the editing process, making it difficult to achieve an optimal solution for the latent representation. To address these issues, we propose a unified gradient-latent purification framework that collects gradient and latent information across different stages to identify effective and concordant update directions. We design a local coordinate system construction method based on feature decomposition, enabling short-term gradients and final-stage latents to be reprojected onto new axes. Then, we employ tailored coefficient regularization terms to effectively aggregate the decomposed information. Additionally, a temporal smoothing axis extension strategy is developed to enhance the temporal coherence of the generated content. Extensive experiments demonstrate that our proposed method outperforms state-of-the-art methods across various editing tasks, delivering superior editing performance. Project page is available in https://unity-in-diversity-editing.github.io.

We present a unified network for simultaneously generating videos and their corresponding entity segmentation and depth maps from text prompts. We utilize colormap to represent entity masks and depth maps, tightly integrating dense prediction with RGB video generation. Introducing dense prediction information improves video generation's consistency and motion smoothness without increasing computational costs. Incorporating learnable task embeddings brings multiple dense prediction tasks into a single model, enhancing flexibility and further boosting performance. We further propose a large-scale dense prediction video dataset Panda-Dense, addressing the issue that existing datasets do not concurrently contain captions, videos, segmentation, or depth maps. Comprehensive experiments demonstrate the high efficiency of our method, surpassing the state-of-the-art in terms of video quality, consistency, and motion smoothness.

This paper presents a unified approach to understanding dynamic scenes from casual videos. Large pretrained vision foundation models, such as vision-language, video depth prediction, motion tracking, and segmentation models, offer promising capabilities. However, training a single model for comprehensive 4D understanding remains challenging. We introduce Uni4D, a multi-stage optimization framework that harnesses multiple pretrained models to advance dynamic 3D modeling, including static/dynamic reconstruction, camera pose estimation, and dense 3D motion tracking. Our results show state-of-the-art performance in dynamic 4D modeling with superior visual quality. Notably, Uni4D requires no retraining or fine-tuning, highlighting the effectiveness of repurposing visual foundation models for 4D understanding. Code and more results are available at: https://davidyao99.github.io/uni4d.

Transmission matrices, mapping the propagation of light from one end of the tissue to the other, form an important mathematical tool in the analysis of tissue scattering and the design of wavefront shaping systems. To understand the relationship between their content and the volumetric structure of the tissue, we wish to fit them with multi-slice models, composed of a set of planar aberrations spaced throughout the volume. The number of layers used in such a model would largely affect the amount of information compression and the ease in which we can use such layered models in a wavefront-shaping system. This work offers a theoretical study of such multi-layered models. We attempt to understand how many layers are required for a good fit, and how does the approximation degrade when a smaller number of such layers is used. We show analytically that transmission matrices can be well fitted with very sparse layers. This leads to optimistic predictions on our ability to use them to design future wavefront shaping systems which can correct tissue aberration over a wide field-of-view.