ICCV 2025 论文列表

Visual understanding is inherently intention-driven--humans selectively focus on different regions of a scene based on their goals. Recent advances in large multimodal models (LMMs) enable flexible expression of such intentions through natural language, allowing queries to guide visual reasoning processes. Frameworks like Visual Chain-of-Thought have demonstrated the benefit of incorporating explicit reasoning steps, where the model predicts a focus region before answering a query. However, existing approaches rely heavily on supervised training with annotated intermediate bounding boxes, which severely limits scalability due to the combinatorial explosion of intention-region pairs. To overcome this limitation, we propose VisRL, the first framework that applies reinforcement learning (RL) to the problem of intention-driven visual perception. VisRL optimizes the entire visual reasoning process using only reward signals. By treating intermediate focus selection as an internal decision optimized through trial-and-error, our method eliminates the need for costly region annotations while aligning more closely with how humans learn to perceive the world. Extensive experiments across multiple benchmarks show that VisRL consistently outperforms strong baselines, demonstrating both its effectiveness and its strong generalization across different LMMs.

Can Multimodal Large Language Models (MLLMs) develop an intuitive number sense similar to humans? Targeting this problem, we introduce Visual Number Benchmark (VisNumBench) to evaluate the number sense abilities of MLLMs across a wide range of visual numerical tasks. VisNumBench consists of about 1,900 multiple-choice question-answer pairs derived from both synthetic and real-world visual data, covering seven visual numerical attributes and four types of visual numerical estimation tasks. Our experiments on VisNumBench led to the following key findings:(i) The 17 MLLMs we tested--including open-source models such as Qwen2.5-VL and InternVL2.5, as well as proprietary models like GPT-4o and Gemini 2.0 Flash--perform significantly below human levels in number sense-related tasks. (ii) Multimodal mathematical models and multimodal chain-of-thought (CoT) models did not exhibit significant improvements in number sense abilities. (iii) Stronger MLLMs with larger parameter sizes and broader general abilities demonstrate modest gains in number sense abilities. We believe VisNumBench will serve as a valuable resource for the research community, encouraging further advancements in enhancing MLLMs' number sense abilities.

This paper introduces VisHall3D, a novel two-stage framework for monocular semantic scene completion that aims to address the issues of feature entanglement and geometric inconsistency prevalent in existing methods. VisHall3D decomposes the scene completion task into two stages: reconstructing the visible regions (vision) and inferring the invisible regions (hallucination). In the first stage, VisFrontierNet, a visibility-aware projection module, is introduced to accurately trace the visual frontier while preserving fine-grained details. In the second stage, OcclusionMAE, a hallucination network, is employed to generate plausible geometries for the invisible regions using a noise injection mechanism. By decoupling scene completion into these two distinct stages, VisHall3D effectively mitigates feature entanglement and geometric inconsistency, leading to significantly improved reconstruction quality.The effectiveness of VisHall3D is validated through extensive experiments on two challenging benchmarks: SemanticKITTI and SSCBench-KITTI-360. VisHall3D achieves state-of-the-art performance, outperforming previous methods by a significant margin and paves the way for more accurate and reliable scene understanding in autonomous driving and other applications.

3D visual grounding aims to identify and localize objects in a 3D space based on textual descriptions. However, existing methods struggle with disentangling targets from anchors in complex multi-anchor queries and resolving inconsistencies in spatial descriptions caused by perspective variations.To tackle these challenges, we propose ViewSRD, a framework that formulates 3D visual grounding as a structured multi-view decomposition process. First, the Simple Relation Decoupling (SRD) module restructures complex multi-anchor queries into a set of targeted single-anchor statements, generating a structured set of perspective-aware descriptions that clarify positional relationships. These decomposed representations serve as the foundation for the Multi-view Textual-Scene Interaction (Multi-TSI) module, which integrates textual and scene features across multiple viewpoints using shared, Cross-modal Consistent View Tokens (CCVTs) to preserve spatial correlations. Finally, a Textual-Scene Reasoning module synthesizes multi-view predictions into a unified and robust 3D visual grounding.Experiments on 3D visual grounding datasets show that ViewSRD significantly outperforms state-of-the-art methods, particularly in complex queries requiring precise spatial differentiation.Code is available at https://github.com/visualjason/ViewSRD

Learning a robust video Variational Autoencoder (VAE) is essential for reducing video redundancy and facilitating efficient video generation. Directly applying image VAEs to individual frames in isolation results in temporal inconsistencies and fails to compress temporal redundancy effectively. Existing works on Video VAEs compress temporal redundancy but struggle to handle videos with large motion effectively. They suffer from issues such as severe image blur and loss of detail in scenarios with large motion. In this paper, we present a powerful video VAE named VideoVAE+ that effectively reconstructs videos with large motion. First, we investigate two architecture choices and propose our simple yet effective architecture with better spatiotemporal joint modeling performance. Second, we propose to leverage the textual information in existing text-to-video datasets and incorporate text guidance during training. The textural guidance is optional during inference. We find that this design enhances the reconstruction quality and preservation of detail. Finally, our models achieve strong performance compared with various baseline approaches in both general videos and large motion videos, demonstrating its effectiveness on the challenging large motion scenarios.

Recognizing subtle similarities and differences among sets of similar activities is central to many real-world applications, including skill acquisition, sports performance evaluation, and anomaly detection. Humans excel at such fine-grained analysis, which requires comprehensive video understanding and cross-video reasoning about action attributes, poses, positions, and emotional states. Yet existing video-based large language models typically address only single-video recognition, leaving their capacity for multi-video reasoning largely unexplored. We introduce VideoSetDiff, a curated dataset designed to test detail-oriented recognition across diverse activities, from subtle action attributes to viewpoint transitions. Our evaluation of current video-based LLMs on VideoSetDiff reveals critical shortcomings, particularly in fine-grained detail recognition and multi-video reasoning. To mitigate these issues, we propose an automatically generated dataset for instruction tuning alongside a novel multi-video recognition framework. While instruction tuning and specialized multi-video reasoning improve performance, all tested models remain far from satisfactory. These findings underscore the need for more robust video-based LLMs capable of handling complex multi-video tasks, enabling diverse real-world applications.

We propose VideoRFSplat, a direct text-to-3D model leveraging a video generation model to generate realistic 3D Gaussian Splatting (3DGS) for unbounded real-world scenes. To generate diverse camera poses and unbounded spatial extent of real-world scenes, while ensuring generalization to arbitrary text prompts, previous methods fine-tune 2D generative models to jointly model camera poses and multi-view images. However, these methods suffer from instability when extending 2D generative models to joint modeling due to the modality gap, which necessitates additional models to stabilize training and inference. In this work, we propose an architecture and a sampling strategy to jointly model multi-view images and camera poses when fine-tuning a video generation model. Our core idea is a dual-stream architecture that attaches a dedicated pose generation model alongside a pre-trained video generation model via communication blocks, generating multi-view images and camera poses through separate streams. This design reduces interference between the pose and image modalities. Additionally, we propose an asynchronous sampling strategy that denoises camera poses faster than multi-view images, allowing rapidly denoised poses to condition multi-view generation, reducing mutual ambiguity and enhancing cross-modal consistency. Trained on multiple large-scale real-world datasets (RealEstate10K, MVImgNet, DL3DV-10K, ACID), VideoRFSplat outperforms existing text-to-3D direct generation methods that heavily depend on post-hoc refinement via score distillation sampling, achieving superior results without such refinement. We propose VideoRFSplat, a direct text-to-3D model leveraging a video generation model to generate realistic 3D Gaussian Splatting (3DGS) for unbounded real-world scenes. To generate diverse camera poses and unbounded spatial extent of real-world scenes, while ensuring generalization to arbitrary text prompts, previous methods fine-tune 2D generative models to jointly model camera poses and multi-view images. However, these methods suffer from instability when extending 2D generative models to joint modeling due to the modality gap, which necessitates additional models to stabilize training and inference. In this work, we propose an architecture and a sampling strategy to jointly model multi-view images and camera poses when fine-tuning a video generation model. Our core idea is a dual-stream architecture that attaches a dedicated pose generation model alongside a pre-trained video generation model via communication blocks, generating multi-view images and camera poses through separate streams. This design reduces interference between the pose and image modalities. Additionally, we propose an asynchronous sampling strategy that denoises camera poses faster than multi-view images, allowing rapidly denoised poses to condition multi-view generation, reducing mutual ambiguity and enhancing cross-modal consistency. Trained on multiple large-scale real-world datasets (RealEstate10K, MVImgNet, DL3DV-10K, ACID), VideoRFSplat outperforms existing text-to-3D direct generation methods that heavily depend on post-hoc refinement via score distillation sampling, achieving superior results without such refinement. Project page: https://gohyojun15.github.io/VideoRFSplat/

We present VideoOrion, a Video Large Language Model (Video-LLM) that explicitly captures the key semantic information in videos--the spatial-temporal dynamics of objects throughout the videos. VideoOrion employs expert vision models to extract object dynamics through a detect-segment-track pipeline, encoding them into a set of object tokens by aggregating spatial-temporal object features. Our method addresses the persistent challenge in Video-LLMs of efficiently compressing high-dimensional video data into semantic tokens that are comprehensible to LLMs. Compared to prior methods which resort to downsampling the original video or aggregating visual tokens using resamplers, leading to information loss and entangled semantics, VideoOrion not only offers a more natural and efficient way to derive compact, disentangled semantic representations but also enables explicit object modeling of video content with minimal computational cost. Moreover, the introduced object tokens naturally allow VideoOrion to accomplish video-based referring tasks. Experimental results show that VideoOrion can learn to make good use of the object tokens, and achieves competitive results on both general video question answering and video-based referring benchmarks.

Understanding hour-long videos with multi-modal large language models (MM-LLMs) enriches the landscape of human-centered AI applications. However, for end-to-end video understanding with LLMs, uniformly sampling video frames results in LLMs being overwhelmed by a vast amount of irrelevant information as video length increases. Existing hierarchical key frame extraction methods improve the accuracy of video understanding but still face two critical challenges. 1) How can the interference of extensive redundant information in long videos be mitigated? 2) How can a model dynamically adapt to complex hierarchical structures while accurately identifying key frames? To address these issues, we propose VideoMiner, which iteratively segments, captions, and clusters long videos, forming a hierarchical tree structure. The proposed VideoMiner progresses from long videos to events to frames while preserving temporal coherence, effectively addressing the first challenge. To precisely locate key frames, we introduce T-GRPO, a tree-based group relative policy optimization in reinforcement learning method that guides the exploration of the VideoMiner. The proposed T-GRPO is specifically designed for tree structures, integrating spatiotemporal information at the event level while being guided by the question, thus solving the second challenge. We achieve superior performance in all long-video understanding tasks and uncover several interesting insights. Our proposed T-GRPO surprisingly incentivizes the model to spontaneously generate a reasoning chain. Additionally, the designed tree growth auxin dynamically adjusts the expansion depth, obtaining accuracy and efficiency gains. The code is publicly available at https://github.com/caoxinye/VideoMiner.

Recent advancements in large-scale video-language models have shown significant potential for real-time planning and detailed interactions. However, their high computational demands and the scarcity of annotated datasets limit their practicality for academic researchers. In this work, we introduce VideoLLaMB, a novel and efficient framework for long video understanding that leverages recurrent memory bridges and temporal memory tokens to enable seamless encoding of entire video sequences with preserved semantic continuity. Central to our approach is a SceneTiling algorithm that segments videos into coherent semantic units, facilitating robust understanding across tasks without requiring additional training. VideoLLaMB achieves state-of-the-art performance, surpassing existing models by 4.2 points on four VideoQA benchmarks and by 2.06 points on egocentric planning tasks. Notably, it maintains strong performance under extreme video length scaling (up to 8x) and excels at fine-grained frame retrieval on our proposed Needle in a Video Haystack (NIAVH) benchmark. With linear GPU memory scaling, VideoLLaMB processes up to 320 frames using a single Nvidia A100 GPU, despite being trained on only 16 frames--offering an unprecedented balance of accuracy, scalability, and cost-effectiveness. This makes it highly accessible and practical for the academic community.

Recent video generation models have shown promising results in producing high-quality video clips lasting several seconds. However, these models face challenges in generating long sequences that convey clear and informative events, limiting their ability to support coherent narrations. In this paper, we present a large-scale cooking video dataset designed to advance long-form narrative generation in the cooking domain. We validate the quality of our proposed dataset in terms of visual fidelity and textual caption accuracy using state-of-the-art Vision-Language Models (VLMs) and video generation models, respectively. We further introduce a Long Narrative Video Director to enhance both visual and semantic coherence in generated videos and emphasize the role of aligning visual embeddings to achieve improved overall video quality. Our method demonstrates substantial improvements in generating visually detailed and semantically aligned keyframes, supported by finetuning techniques that integrate text and image embeddings within the video generation process. Codes and data will be made publicly available.

Existing approaches to drone visual geo-localization predominantly adopt the image-based setting, where a single drone-view snapshot is matched with images from other platforms. Such task formulation, however, underutilizes the inherent video output of the drone and is sensitive to occlusions and viewpoint disparity. To address these limitations, we formulate a new video-based drone geo-localization task and propose the Video2BEV paradigm. This paradigm transforms the video into a Bird's Eye View (BEV), simplifying the subsequent inter-platform matching process. In particular, we employ Gaussian Splatting to reconstruct a 3D scene and obtain the BEV projection. Different from the existing transform methods, e.g., polar transform, our BEVs preserve more fine-grained details without significant distortion. To facilitate the discriminative intra-platform representation learning, our Video2BEV paradigm also incorporates a diffusion-based module for generating hard negative samples. To validate our approach, we introduce UniV, a new video-based geo-localization dataset that extends the image-based University-1652 dataset. UniV features flight paths at 30^\circ and 45^\circ elevation angles with increased frame rates of up to 10 frames per second (FPS). Extensive experiments on the UniV dataset show that our Video2BEV paradigm achieves competitive recall rates and outperforms conventional video-based methods. Compared to other competitive methods, our proposed approach exhibits robustness at lower elevations with more occlusions. The code is available at: https://github.com/HaoDot/Video2BEV-Open.

With the scale capability of increasing training data, model size, and computational cost, video generation has achieved impressive results in digital creation, enabling users to express creativity across various domains. Recently, researchers in Large Language Models (LLMs) have expanded the scaling to test-time, which can significantly improve LLM performance by using more inference-time computation. Instead of scaling up video foundation models through expensive training costs, we explore the power of Test-Time Scaling (TTS) in video generation, aiming to answer the question: if a video generation model is allowed to use non-trivial amount of inference-time compute, how much can it improve generation quality given a challenging text prompt. In this work, we reinterpret the test-time scaling of video generation as a searching problem to sample better trajectories from Gaussian noise space to the target video distribution. Specifically, we build the search space with test-time verifiers to provide feedback and heuristic algorithms to guide searching process. Given a text prompt, we first explore an intuitive linear search strategy by increasing noise candidates at inference time. As full-step denoising all frames simultaneously requires heavy test-time computation costs, we further design a more efficient TTS method for video generation called Tree-of-Frames (ToF) that adaptively expands and prunes video branches in an autoregressive manner. Extensive experiments on text-conditioned video generation benchmarks demonstrate that increasing test-time compute consistently leads to significant improvements in the quality of videos.

We present Video Motion Graphs, a system designed to generate realistic human motion videos. Using a reference video and conditional signals such as music or motion tags, the system synthesizes new videos by first retrieving video clips with gestures matching the conditions and then generating interpolation frames to seamlessly connect clip boundaries. The core of our approach is HMInterp, a robust Video Frame Interpolation (VFI) model that enables seamless interpolation of discontinuous frames, even for complex motion scenarios like dancing. HMInterp i) employs a dual-branch interpolation approach, combining a Motion Diffusion Model for human skeleton motion interpolation with a diffusion-based video frame interpolation model for final frame generation. ii) adopts condition progressive training to effectively leverage identity strong and weak conditions, such as images and pose. These designs ensure both high video texture quality and accurate motion trajectory. Our Video Motion Graphs outperforms existing generative- and retrieval-based methods for human motion video generation. Our codes and pretrained models are public available.

Video Individual Counting (VIC) has received increasing attention for its importance in intelligent video surveillance. Existing works are limited in two aspects, i.e., dataset and method. Previous datasets are captured with fixed or rarely moving cameras with relatively sparse individuals, restricting evaluation for a highly varying view and time in crowded scenes. Existing methods rely on localization followed by association or classification, which struggle under dense and dynamic conditions due to inaccurate localization of small targets. To address these issues, we introduce the MovingDroneCrowd Dataset, featuring videos captured by fast-moving drones in crowded scenes under diverse illuminations, shooting heights and angles. We further propose a Shared Density map-guided Network (SDNet) using a Depth-wise Cross-Frame Attention (DCFA) module to directly estimate shared density maps between consecutive frames, from which the inflow and outflow density maps are derived by subtracting the shared density maps from the global density maps. The inflow density maps across frames are summed up to obtain the number of unique pedestrians in a video. Experiments on our datasets and publicly available ones show the superiority of our method over the state of the arts in highly dynamic and complex crowded scenes. Our dataset and codes have been released publicly.

Different from color correction and transfer, color grading involves adjusting colors for artistic or storytelling purposes in a video, which is used to establish a specific look or mood. However, due to the complexity of the process and the need for specialized editing skills, video color grading remains primarily the domain of professional colorists. In this paper, we present a reference based video color grading framework. Our key idea is explicitly generating a look-up table (LUT) for color attribute alignment between reference scenes and input video via a diffusion model. As a training objective, we enforce that high-level features of the reference scenes like look, mood, and emotion should be similar to that of the input video. Our LUT-based approach allows for color grading without any loss of structural details in the whole video frames as well as achieving fast inference. We further build a pipeline to incorporate a user-preference via text prompts for low-level feature enhancement such as contrast and brightness, etc. Experimental results, including extensive user studies, demonstrate the effectiveness of our approach for video color grading.

Given a natural language query, temporal video grounding aims to localize the described temporal moment in an untrimmed video. A major challenge of this task is its heavy dependence on labor-intensive annotations for training. Unlike existing works that directly train models on manually curated data, we propose a novel paradigm to reduce annotation costs: pretraining the model on unlabeled, real-world videos. To support this, we introduce Temporal Video Grounding Pretraining (Vid-Group), a large-scale dataset collected in a scalable manner with minimal human intervention, consisting of over 50K videos captured in the wild and 200K pseudo annotations. Direct pretraining on these imperfect pseudo annotations, however, presents significant challenges, including mismatched sentence-video pairs and imprecise temporal boundaries. To address these issues, we propose the ReCorrect algorithm, which comprises two main phases: semantics-guided refinement and memory-consensus correction. The semantics-guided refinement enhances the pseudo labels by leveraging semantic similarity with video frames to clean out unpaired data and make initial adjustments to temporal boundaries. In the following memory-consensus correction phase, a memory bank tracks the model predictions, progressively correcting the temporal boundaries based on consensus within the memory. Comprehensive experiments demonstrate ReCorrect's strong generalization abilities across multiple downstream settings. The code, dataset, and pretrained models are available at https://github.com/baopj/Vid-Group.

Vision foundation models (VFMs) have demonstrated remarkable performance across a wide range of downstream tasks. While several VFM adapters have shown promising results by leveraging the prior knowledge of VFMs, we identify two inefficiencies in these approaches. First, the interaction between convolutional neural network (CNN) and VFM backbone triggers early layer gradient backpropagation. Second, existing methods require tuning all components, adding complexity. Besides, these adapters alter VFM features, underutilizing the prior knowledge. To tackle these challenges, we propose a new approach called ViT-Split, based on a key observation: the layers of several VFMs, like DINOv2, can be divided into two distinct components: an extractor for learning low-level features and an adapter for learning task-specific features. Leveraging this insight, we eliminate the CNN branch and introduce two heads, task head and prior head, to the frozen VFM. The task head is designed to learn task-specific features, mitigating the early gradient propagation issue. The prior head is used to leverage the multi-scale prior features from the frozen VFM, reducing tuning parameters and overfitting. Extensive experiments on various tasks (e.g., segmentation, detection, depth estimation, and visual question answering) validate the effectiveness and efficiency of ViT-Split. Specifically, ViT-Split reduces training time up to 4xwhile achieving comparable or even better results on ADE20K, compared to other VFM adapters.