CVPR 2025 论文列表

In this work, we systematically study music generation conditioned solely on the video. First, we present a large-scale dataset by collecting 360K video-music pairs, including various genres such as movie trailers, advertisements, and documentaries. Furthermore, we propose VidMuse, a simple framework for generating music aligned with video inputs. VidMuse stands out by producing high-fidelity music that is both acoustically and semantically aligned with the video. By incorporating local and global visual cues, VidMuse enables the creation of coherent music tracks that consistently match the video content through Long-Short-Term modeling. Through extensive experiments, VidMuse outperforms existing models in terms of audio quality, diversity, and audio-visual alignment. The code and datasets are available at https://vidmuse.github.io/

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/

Future robots are envisioned as versatile systems capable of performing a variety of household tasks. The big question remains, how can we bridge the embodiment gap while minimizing physical robot learning, which fundamentally does not scale well. We argue that learning from in-the-wild human videos offers a promising solution for robotic manipulation tasks, as vast amounts of relevant data already exist on the internet. In this work, we present VidBot, a framework enabling zero-shot robotic manipulation using learned 3D affordance from in-the-wild monocular RGB-only human videos. VidBot leverages a pipeline to extract explicit representations from them, namely 3D hand trajectories from videos, combining a depth foundation model with structure-from-motion techniques to reconstruct temporally consistent, metric-scale 3D affordance representations agnostic to embodiments. We introduce a coarse-to-fine affordance learning model that first identifies coarse actions from the pixel space and then generates fine-grained interaction trajectories with a diffusion model, conditioned on coarse actions and guided by test-time constraints for context-aware interaction planning, enabling substantial generalization to novel scenes and embodiments. Extensive experiments demonstrate the efficacy of VidBot, which significantly outperforms counterparts across 13 manipulation tasks in zero-shot settings and can be seamlessly deployed across robot systems in real-world environments. VidBot paves the way for leveraging everyday human videos to make robot learning more scalable.

Sim-to-real gap has long posed a significant challenge for robot learning in simulation, preventing the deployment of learned models in the real world. Previous work has primarily focused on domain randomization and system identification to mitigate this gap. However, these methods are often limited by the realism of the simulation and graphics engines. In this work, we propose Vid2Sim, a novel framework that effectively bridges the sim2real gap through a scalable and cost-efficient real2sim pipeline for neural 3D scene reconstruction and simulation. Given a monocular video as input, Vid2Sim can generate photo-realistic and physically interactable 3D simulation environments to enable the reinforcement learning of visual navigation agents in complex urban environments. Extensive experiments demonstrate that Vid2Sim significantly improves the performance of urban navigation in the digital twins and real world by 31.2% and 68.3% in success rate compared with agents trained with prior simulation methods. Code and data will be made publicly available.

Faithfully reconstructing textured shapes and physical properties from videos presents an intriguing yet challenging problem. Significant efforts have been dedicated to advancing such a system identification problem in this area. Previous methods often rely on heavy optimization pipelines with a differentiable simulator and renderer to estimate physical parameters. However, these approaches frequently necessitate extensive hyperparameter tuning for each scene and involve a costly optimization process, which limits both their practicality and generalizability. In this work, we propose a novel framework, Vid2Sim, a generalizable video-based approach for recovering geometry and physical properties through a mesh-free reduced simulation based on Linear Blend Skinning (LBS), offering high computational efficiency and versatile representation capability. Specifically, Vid2Sim first reconstructs the observed configuration of the physical system from video using a feed-forward neural network trained to capture physical world knowledge. A lightweight optimization pipeline then refines the estimated appearance, geometry, and physical properties to closely align with video observations within just a few minutes. Additionally, after the reconstruction, Vid2Sim enables high-quality, mesh-free simulation with high efficiency. Extensive experiments demonstrate that our method achieves superior accuracy and efficiency in reconstructing geometry and physical properties from video data.

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.

Programming based approaches to reasoning tasks have substantially expanded the types of questions models can answer about visual scenes.Yet on benchmark visual reasoning data, when answering correctly, such models produce incorrect programs 33% of the time. These models are often right for the wrong reasons and risk unexpected failures on new data. Unit tests play a foundational role in ensuring code correctness and could be used to repair such failures. We propose Visual Unit Testing (ViUniT), a framework to improve the reliability of visual programs by automatically generating unit tests. In our framework, a unit test is represented as a novel image and answer meant to verify logical correctness of a program produced for a given query. Our method leverages a language model to create unit tests in the form of image descriptions and expected answers and image synthesis to produce corresponding images. We conduct a comprehensive analysis of what constitutes an effective visual unit test suite, exploring unit test generation, sampling strategies, image generation methods, and varying the number of programs and unit tests. Additionally, we introduce four applications of visual unit tests: best program selection, answer refusal, re-prompting, and unsupervised reward formulations for reinforcement learning. Experiments with two models across three datasets in visual question answering and image-text matching demonstrate that ViUniT improves model performance by 11.4%. Notably, it enables 7B open-source models to outperform gpt-4o-mini by an average of 7.7% and reduces the occurrence of programs that are correct for the wrong reasons by 40%.

The sparsity of point clouds and inadequacy of semantic information pose challenges to current LiDAR-only 3D object detection methods. Recent methods alleviate these challenges by converting RGB images into virtual points via depth completion to be fused with LiDAR points. Although these methods have shown outstanding results, they often introduce significant computation overhead due to the high density of virtual points and noise due to inaccurate depth completion. Besides, they do not thoroughly leverage semantic information from images. In this work, we propose the virtual key instance enhanced network (ViKIENet), a highly efficient and effective multi-modal feature fusion framework that fuses the features of virtual key instances (VKIs) and LiDAR points through multiple stages. Our contributions include three main components: semantic key instance selection (SKIS), virtual-instance-focused fusion (VIFF), and virtual-instance-to-real attention (VIRA). We also propose the extended version ViKIENet-R with VIFF-R which includes rotationally equivariant features. Experiment results show that ViKIENet and ViKIENet-R achieve significant improvements in detection performance on the KITTI, JRDB, and nuScenes datasets compared to existing works. On the KITTI dataset, ViKIENet and ViKIENet-R operate at 22.7 and 15.0 FPS, respectively. As of CVPR submission (Nov. 15th, 2024), ViKIENet ranks first on the car detection and orientation estimation leaderboard, while ViKIENet-R ranks second (compared with officially published papers) on the 3D car detection leaderboard.

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/

Recent large-scale text-to-image diffusion models generate photorealistic images but often struggle to accurately depict interactions between humans and objects due to their limited ability to differentiate various interaction words.In this work, we propose VerbDiff to address the challenge of capturing nuanced interactions within text-to-image diffusion models. VerbDiff is a novel text-to-image generation model that weakens the bias between interaction words and objects, enhancing the understanding of interactions. Specifically, we disentangle various interaction words from frequency-based anchor words and leverage localized interaction regions from generated images to help the model better capture semantics in distinctive words without extra conditions. Our approach enables the model to accurately understand the intended interaction between humans and objects, producing high-quality images with accurate interactions aligned with specified verbs. Extensive experiments on the HICO-DET dataset demonstrate the effectiveness of our method compared to previous approaches.

Angiography imaging is a medical imaging technique that enhances the visibility of blood vessels within the body by using contrast agents. Angiographic images can effectively assist in the diagnosis of vascular diseases. However, contrast agents may bring extra radiation exposure which is harmful to patients with health risks. To mitigate these concerns, in this paper, we aim to automatically generate angiography from non-angiographic inputs, by leveraging and enhancing the inherent physical properties of vascular structures. Previous methods relying on 2D slice-based angiography synthesis struggle with maintaining continuity in 3D vascular structures and exhibit limited effectiveness across different imaging modalities. We propose VasTSD, a 3D vascular tree-state space diffusion model to synthesize angiography from 3D non-angiographic volumes, with a novel state space serialization approach that dynamically constructs vascular tree topologies, integrating these with a diffusion-based generative model to ensure the generation of anatomically continuous vasculature in 3D volumes. A pre-trained vision embedder is employed to construct vascular state space representations, enabling consistent modeling of vascular structures across multiple modalities. Extensive experiments on various angiographic datasets demonstrate the superiority of VasTSD over prior works, achieving enhanced continuity of blood vessels in synthesized angiographic synthesis in multiple modalities and anatomical regions.

The proliferation of multi-modal generative models has introduced new privacy and security challenges, especially due to the risks of memorization and unintentional disclosure of sensitive information. This paper focuses on the vulnerability of multi-modal image captioning models to membership inference attacks (MIAs). These models, which synthesize textual descriptions from visual content, could inadvertently reveal personal or proprietary data embedded in their training datasets. We explore the feasibility of MIAs in the context of such models. Specifically, our approach leverages a variance-based strategy tailored for image captioning models, utilizing only image data without knowing the corresponding caption. We introduce the means-of-variance threshold attack (MVTA) and confidence-based weakly supervised attack (C-WSA) based on the metric, means-of-variance (MV), to assess variability among vector embeddings. Our experiments demonstrate that these models are susceptible to MIAs, indicating substantial privacy risks. The effectiveness of our methods is validated through rigorous evaluations on these real-world models, confirming the practical implications of our findings.

We present V2Dial - a novel expert-based model specifically geared towards simultaneously handling image and video input data for multimodal conversational tasks. Current multimodal models primarily focus on simpler tasks (e.g., VQA, VideoQA, video-text retrieval) and often neglect the more challenging conversational counterparts, such as video and visual/image dialog. Moreover, works on both conversational tasks evolved separately from each other despite their apparent similarities limiting their applicability potential. To this end, we propose to unify both tasks using a single model that for the first time jointly learns the spatial and temporal features of images and videos by routing them through dedicated experts and aligns them using matching and contrastive learning techniques. Furthermore, we systemically study the domain shift between the two tasks by investigating whether and to what extent these seemingly related tasks can mutually benefit from their respective training data. Extensive evaluations on the widely used video and visual dialog datasets of AVSD and VisDial show that our model achieves new state-of-the-art results across four benchmarks both in zero-shot and fine-tuning settings.

Although diffusion-based image virtual try-on has made considerable progress, emerging approaches still struggle to effectively address the issue of hand occlusion (i.e., clothing regions occluded by the hand part), leading to a notable degradation of the try-on performance. To tackle this issue widely existing in real-world scenarios, we propose VTON-HandFit, leveraging the power of hand priors to reconstruct the appearance and structure for hand occlusion cases. Firstly, we tailor a Hand-Pose Aggregation Net using the ControlNet-based structure explicitly and adaptively encoding the global hand and pose priors. Besides, to fully exploit the hand-related structure and appearance information, we propose Hand-feature Disentanglement Embedding module to disentangle the hand priors into the hand structure-parametric and visual-appearance features, and customize a masked cross attention for further decoupled feature embedding. Lastly, we customize a hand-canny constraint loss to better learn the structure edge knowledge from the hand template of model image. VTON-HandFit outperforms the baselines in qualitative and quantitative evaluations on the public dataset and our self-collected hand-occlusion Handfit-3K dataset particularly for the arbitrary hand pose occlusion cases in real-world scenarios.

Virtual Try-On (VTON) is a transformative technology in e-commerce and fashion design, enabling realistic digital visualization of clothing on individuals. In this work, we propose VTON 360, a novel 3D VTON method that addresses the open challenge of achieving high-fidelity VTON that supports any-view rendering. Specifically, we leverage the equivalence between a 3D model and its rendered multi-view 2D images, and reformulate 3D VTON as an extension of 2D VTON that ensures 3D consistent results across multiple views.To achieve this, we extend 2D VTON models to include multi-view garments and clothing-agnostic human body images as input, and propose several novel techniques to enhance them, including: i) a pseudo-3D pose representation using normal maps derived from the SMPL-X 3D human model, ii) a multi-view spatial attention mechanism that models the correlations between features from different viewing angles, and iii) a multi-view CLIP embedding that enhances the garment CLIP features used in 2D VTON with camera information. Extensive experiments on large-scale real datasets and clothing images from e-commerce platforms demonstrate the effectiveness of our approach.

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.

The recent surge in high-quality visual instruction tuning samples from closed-source vision-language models (VLMs) such as GPT-4V has accelerated the release of open-source VLMs across various model sizes. However, scaling VLMs to improve performance using larger models brings significant computational challenges, especially for deployment on resource-constrained devices like mobile platforms and robots. To address this, we propose VLsI: Verbalized Layers-to-Interactions, a new VLM family in 2B and 7B model sizes, which prioritizes efficiency without compromising accuracy. VLsI leverages a unique, layer-wise distillation process, introducing intermediate "verbalizers" that map features from each layer to natural language space, allowing smaller VLMs to flexibly align with the reasoning processes of larger VLMs. This approach mitigates the training instability often encountered in output imitation and goes beyond typical final-layer tuning by aligning the small VLMs' layer-wise progression with that of the large ones. We validate VLsI across ten challenging vision-language benchmarks, achieving notable performance gains (11.0% for 2B and 17.4% for 7B) over GPT-4V without the need for model scaling, merging, or architectural changes.