CVPR 2026 Papers — Page 39
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
Unlocking Pre-trained Weights: Parameter Inheritance for Zero-Shot Initialization
Jiaze Xu (Southeast University), Xin Geng (Southeast University)
Knowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningImage
🎯 What it does: Propose Parameter Inheritance HyperNetwork (PITH), which uses a learnable projection matrix to directly map weights from publicly pre-trained models to the target model, achieving zero-shot initialization;
Unlocking Strong Supervision: A Data-Centric Study of General-Purpose Audio Pre-Training Methods
Xuanru Zhou (Zhejiang University), Dong Yu (Tencent AI Lab)
Representation LearningData-Centric LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningTextMultimodalityAudio
🎯 What it does: Explored large-scale audio pre-training, proposing a strongly supervised data center pipeline, constructing a unified tag system (UTS) and generating high-quality captions, followed by a systematic comparison of multiple pre-training objectives on the same dataset.
Unlocking the Power of Critical Factors for 3D Visual Geometry Estimation
Guangkai Xu (Zhejiang University), Chunhua Shen (Zhejiang University)
Pose EstimationDepth EstimationTransformerVideoPoint Cloud
🎯 What it does: This paper proposes the CARVE framework, which improves the accuracy and consistency of single-frame and multi-frame 3D visual geometry estimation by leveraging a more diverse dataset, removing gradient and adaptive confidence loss, introducing geometric consistency loss, and enhancing high-resolution feature fusion.
Unlocking Token Rewards via Training-Free Reward Attribution
Sitong Wu (Chinese University of Hong Kong), Jiaya Jia (Chinese University of Hong Kong)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextMultimodality
🎯 What it does: Proposed a training-agnostic, gradient-based first-order Taylor approximation process reward model (PRM) vectorization, decomposing macro process rewards into fine-grained token-level rewards to achieve precise credit assignment;
Unpaired Image Deraining Using Reward-Guided Self-Reinforcement Strategy
Yinghao Chen (National University of Defense Technology), Yaowen Fu (National University of Defense Technology)
RestorationConvolutional Neural NetworkReinforcement LearningVision Language ModelGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes an unsupervised single-image deraining framework called RGSUD, which utilizes intermediate high-quality deraining results as a reward to guide model learning;
Unposed-to-3D: Learning Simulation-Ready Vehicles from Real-World Images
Hongyuan Liu (University of Science and Technology Beijing), Huimin Ma (University of Science and Technology Beijing)
GenerationData SynthesisDiffusion modelContrastive LearningGaussian SplattingImageMesh
🎯 What it does: A fully image-supervised framework was constructed, capable of directly generating 3D vehicle models for simulation from real-world driving images without pose annotations.
UnReflectAnything: RGB-Only Highlight Removal by Rendering Synthetic Specular Supervision
Alberto Rota (Politecnico di Milano), Benjamin Busam (Technical University of Munich)
RestorationPose EstimationTransformerImageBiomedical Data
🎯 What it does: Propose a reflection removal method called UnReflectAnything that utilizes single-frame RGB images to eliminate specular highlights and restore diffuse details in natural scenes and surgical endoscopy images.
Unsafe2Safe: Controllable Image Anonymization for Downstream Utility
Minh Dinh (Dartmouth College), SouYoung Jin (Dartmouth College)
Safty and PrivacyLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImage
🎯 What it does: Proposes an end-to-end automated pipeline called Unsafe2Safe, which detects privacy risks in images and rewrites sensitive regions through multimodal reasoning and text-guided diffusion editors, generating anonymized images that preserve semantic structure while removing identity information.
Unstitching the Chimera: Frame-Level Risk and Train-Free Mitigation for Video Hallucination
Songyuan Yang (National University of Defense Technology), Huibin Tan (National University of Defense Technology)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerVision Language ModelVideoTextBenchmark
🎯 What it does: This paper addresses the 'Chimera hallucination' problem in video multimodal large language models by proposing a unidirectional no-reference risk assessment, CH-Risk, and a two-stage inference correction method, CH-M, based on this risk. The goal is to reduce narrative errors caused by cross-segment concatenation and stage mismatch through frame-level attention rerouting and residual token calibration.
Unsupervised 3d Motion Estimation Using Event Camera
Han Han (University of Science and Technology of China), Zheng-jun Zha (University of Science and Technology of China)
Pose EstimationAutonomous DrivingConvolutional Neural NetworkContrastive LearningOptical Flow
🎯 What it does: This paper proposes an unsupervised event camera 3D motion estimation framework that jointly predicts optical flow and motion (MID), refines MID using horizontal/vertical expansion information from event projection, and finally optimizes the entire model through a contrast maximization objective.
Unsupervised Monocular 3D Keypoint Discovery from Multi-View Diffusion Priors
Subin Jeon (Yonsei University), Seon Joo Kim (Yonsei University)
GenerationPose EstimationDiffusion modelImage
🎯 What it does: Leveraging a pre-trained multi-view diffusion model, KeyDiff3D unsupervisedly discovers and predicts 3D keypoints from a single image, enabling animation manipulation of generated 3D objects based on these keypoints.
Unsupervised Multi-agent and Single-agent Perception from Cooperative Views
Haochen Yang (Cleveland State University), Hongkai Yu (Cleveland State University)
Object DetectionAutonomous DrivingConvolutional Neural NetworkContrastive LearningPoint Cloud
🎯 What it does: Proposed the UMS framework under no manual annotation conditions, jointly solving the 3D object detection problem of multi-vehicle collaboration and single-vehicle perception;
Unsupervised Multi-Scale Segmentation of 3D Subcellular World with Stable Diffusion Foundation Model
Mostofa Rafid Uddin (Carnegie Mellon University), Min Xu (Carnegie Mellon University)
SegmentationConvolutional Neural NetworkDiffusion modelBiomedical DataComputed Tomography
🎯 What it does: This paper proposes a fully unsupervised multi-scale subcellular structure segmentation method, which utilizes the Stable Diffusion pre-trained model to extract features and generates segmentation masks through spectral clustering and heuristic aggregation. Subsequently, CellPose is used to separate membranes and macromolecules, and finally, UNet and DeepETPicker are trained to achieve membrane segmentation and macromolecule localization.
UPLiFT: Efficient Pixel-Dense Feature Upsampling with Local Attenders
Matthew Walmer (University of Maryland), Abhinav Shrivastava (University of Maryland)
SegmentationGenerationDepth EstimationSuper ResolutionConvolutional Neural NetworkAuto EncoderImageTextMultimodality
🎯 What it does: Proposed an efficient pixel-level feature upsampling method called UPLiFT, which maps low-resolution visual backbones to high-resolution features using iterative convolutions and a local attender.
Upsample Anything: A Simple and Hard to Beat Baseline for Feature Upsampling
Minseok Seo (KAIST), Changick Kim (KAIST)
SegmentationDepth EstimationSuper ResolutionGaussian SplattingImage
🎯 What it does: Learn pixel-level Gaussian kernels through test-time optimization using a single image, upscaling low-resolution features from Vision Foundation Models to high-resolution without any training.
Urban-GS: A Unified 3D Gaussian Splatting Framework for Compact and High-Fidelity Aerial-to-Street Reconstruction
Meng Wang (Beihang University), Yue Qi (Beihang University)
GenerationOptimizationGaussian SplattingImageMultimodality
🎯 What it does: Propose the Urban-GS framework, which utilizes 3D Gaussian Splatting to achieve unified modeling and high-quality rendering of urban scenes from aerial and ground views.
URICA: A Uniformity Region Affine Identifier Capture Algorithm for Arbitrary Region Retrieval in Pathology Images
Ri Su (Hong Kong University of Science and Technology (Guangzhou)), Lei Chen (Hong Kong University of Science and Technology (Guangzhou))
RetrievalTransformerImageBiomedical Data
🎯 What it does: Proposed URICA, a method for retrieving arbitrary scale and orientation WSI regions by leveraging semantic tessellation and affine identifiers.
URScenes: A Multi-scenario Dataset for Unstructured Road Environments
Runsen Liu (Beihang University), Wenwen Luo (Beihang University)
Autonomous DrivingImageMultimodalityPoint CloudBenchmark
🎯 What it does: This paper proposes the URScenes dataset, a multi-scenario and multi-modal perception dataset specifically constructed for unstructured road environments, supporting 3D object detection, multi-object tracking, and occupancy prediction.
UST-Hand: An Uncertainty-aware Spatiotemporal Point Cloud Interaction Network for 3D Self-supervised Hand Pose Estimation
Tianhao Han (Shanghai Jiao Tong University), Erwei Yin (Shanghai Jiao Tong University)
Pose EstimationConvolutional Neural NetworkTransformerFlow-based ModelPoint Cloud
🎯 What it does: Developed a self-supervised multi-view 3D hand pose estimation framework named UST-Hand, utilizing uncertainty modeling and spatiotemporal point cloud interaction to achieve high-precision reconstruction.
UTPTrack: Towards Simple and Unified Token Pruning for Visual Tracking
Hao Wu (Eastern Institute of Technology), Xiaoyu Shen (Eastern Institute of Technology)
Object TrackingComputational EfficiencyTransformerVision Language ModelImageVideoTextMultimodality
🎯 What it does: Proposes UTPTrack, a unified token pruning framework that simultaneously compresses the search region, dynamic template, and static template in state-of-the-art Transformer visual tracking, achieving efficient real-time tracking.
UVU: Improving Multimodal Understanding via Vision-Language Unified Autoregressive Paradigm
Zhehan Kan (Tsinghua University), Xing Sun (Tencent Youtu Lab)
TransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Propose the UVU framework, which directly introduces pixel-level visual supervision during the pre-training phase. By leveraging continuous visual encoding and iterative hierarchical clustering, a 200k pixel-level visual codebook is constructed, achieving unified autoregressive generation of vision and language.
UZ3DVG: Unaided Zero-Shot 3D Visual Grounding with Generated Language Conditions
Wenbin Tan (Xiamen University), Yanyun Qu (Xiamen University)
Object DetectionKnowledge DistillationConvolutional Neural NetworkVision Language ModelTextPoint CloudChain-of-Thought
🎯 What it does: This paper proposes an Unsupervised Zero-shot 3D Visual Grounding (UZ3DVG) framework capable of achieving target localization during inference using only 3D point clouds, without relying on 2D images or external language models;
V-Attack: Targeting Disentangled Value Features for Controllable Adversarial Attacks on LVLMs
Sen Nie (State Key Laboratory of AI Safety Institute of Computing Technology Chinese Academy of Sciences), Xilin Chen (State Key Laboratory of AI Safety Institute of Computing Technology Chinese Academy of Sciences)
Adversarial AttackTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Proposes a controllable adversarial attack method called V-Attack based on the internal value features of Transformer.
V-DPM: 4D Video Reconstruction with Dynamic Point Maps
Edgar Sucar (University of Oxford), Andrea Vedaldi (University of Oxford)
GenerationPose EstimationDepth EstimationTransformerDiffusion modelScore-based ModelSimultaneous Localization and MappingVideoPoint Cloud
🎯 What it does: Propose a method for 4D reconstruction from multi-frame videos in a single pass, extending Dynamic Point Maps (DPM) to multi-view video formats.
V-RGBX: Video Editing with Accurate Controls over Intrinsic Properties
Ye Fang (Fudan University), Tuanfeng Yang Wang (Adobe Research)
Image TranslationGenerationTransformerDiffusion modelAuto EncoderContrastive LearningVideo
🎯 What it does: Propose V-RGBX, an end-to-end video editing framework based on intrinsic properties (color, normal, material, irradiance), capable of editing keyframes and consistently propagating these edits across the temporal domain.
V^2-SAM: Marrying SAM2 with Multi-Prompt Experts for Cross-View Object Correspondence
Jiancheng Pan (Fudan University), Yuqian Fu (Fudan University)
SegmentationConvolutional Neural NetworkTransformerPrompt EngineeringMixture of ExpertsContrastive LearningImageVideoBenchmark
🎯 What it does: This paper proposes V-SAM, a unified framework that migrates SAM2 to cross-view object correspondence tasks, leveraging geometric Anchor and visual Visual prompt generators, and achieving precise segmentation through multi-expert and cyclic consistency selectors.
V2U4Real: A Real-world Large-scale Dataset for Vehicle-to-UAV Cooperative Perception
Weijia Li (Xiamen University), Chenglu Wen (Xiamen University)
Object DetectionObject TrackingAutonomous DrivingConvolutional Neural NetworkImageMultimodalityPoint CloudBenchmark
🎯 What it does: Proposes the V2U4Real dataset and benchmark, which records real-world large-scale multimodal data from vehicle and drone collaborative perception, supporting single-machine and collaborative 3D detection and tracking tasks.
VA-p: Variational Policy Alignment for Pixel-Aware Autoregressive Generation
Xinyao Liao (Huazhong University Of Science And Technology), Angela Yao (National University Of Singapore)
GenerationTransformerReinforcement LearningImageTextMultimodality
🎯 What it does: Lightweight post-training of existing autoregressive image generation models, aligning the generator directly in pixel space using variational inference and reinforcement learning to enhance image quality and diversity.
VABench: A Comprehensive Benchmark for Audio-Video Generation
Daili Hua (Peking University), Wentao Zhang (Peking University)
GenerationTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmarkAudio
🎯 What it does: This paper proposes VABench, a comprehensive evaluation framework for synchronized audio-visual generation, covering three tasks: text-to-audio-visual, image-to-audio-visual, and 3D audio-visual.
VAD-GS: Visibility-Aware Densification for 3D Gaussian Splatting in Dynamic Urban Scenes
Yikang Zhang (Tongji University), Rui Fan (Tongji University)
Autonomous DrivingGaussian SplattingImagePoint Cloud
🎯 What it does: Proposes a visualization-aware densification framework called VAD-GS, which utilizes voxel visibility reasoning, perspective diversification selection, and multi-view stereo reconstruction to complete missing geometric information in dynamic urban scenes and achieve high-quality real-time rendering.
Vanast: Virtual Try-On with Human Image Animation via Synthetic Triplet Supervision
Hyunsoo Cha, Hanbyul Joo
Image TranslationImage HarmonizationSegmentationGenerationData SynthesisPose EstimationTransformerVision Language ModelDiffusion modelImageVideoTextMultimodality
🎯 What it does: Proposed Vanast, a single-stage unified framework capable of directly generating clothing transfer animation videos from a single portrait, a set of clothing images, and a pose video.
VAR RL Done Right: Tackling Asynchronous Policy Conflicts in Visual Autoregressive Generation
Shikun Sun (Tsinghua University), Xinglong Wu (ByteDance)
GenerationTransformerReinforcement LearningDiffusion modelImageText
🎯 What it does: Diagnose the asynchronous policy conflict problem in visual autoregressive (VAR) models within reinforcement learning and propose an improved RL framework, significantly enhancing the quality of text rendering and human preference scores.
Variation-aware Vision Token Dropping for Faster Large Vision-Language Models
Junjie Chen (Sichuan University), Honggang Chen (Shanghai Jiao Tong University)
CompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelImageVideoBenchmark
🎯 What it does: Proposes a dynamic pruning method (Variation-aware Vision Token Dropping, V Drop 2) based on the variation degree of visual tokens within LLMs, discarding layer-by-layer the tokens with the least variation during inference to significantly reduce the number of visual tokens and improve inference speed.
Variational Graph-based Normal Integration
Lixiong Chen (University of Oxford), Imari Sato (National Institute of Informatics)
Depth EstimationPoint CloudMesh
🎯 What it does: Proposed a graph-based variational inference framework for depth recovery from surface normals, supporting both regular grids and scattered point data representations.
VarSplat: Uncertainty-aware 3D Gaussian Splatting for Robust RGB-D SLAM
Anh Thuan Tran (George Mason University), Jana Kosecka (George Mason University)
Gaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: Proposes VarSplat, an uncertainty-aware RGB-D SLAM system built upon 3D Gaussian Splatting (3DGS), capable of online learning the appearance variance of each Gaussian and rendering pixel uncertainty maps;
VAST: Video Ability-Stratified Taxonomy for Data-Efficient Video Reasoning
Zhongan Wang (CCAI, Zhejiang University), Hehe Fan (CCAI, Zhejiang University)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelReinforcement LearningVision Language ModelVideoMultimodality
🎯 What it does: Built a VAST cognitive hierarchy classification system based on Piaget's cognitive development theory, and designed a 15K-scale ability-tiered training set VAST-15K and the corresponding evaluation benchmark VAST-Bench; proposed the VideoVAST reinforcement learning framework, using consistency rewards to improve video reasoning quality.
VCP-Attack: Visual-Contrastive Projection for Transferable Black-Box Targeted Attacks on Large Vision-Language Models
Jiawei Zhao (Southeast University), Yining Hu (Southeast University)
Adversarial AttackVision Language ModelContrastive LearningImageText
🎯 What it does: Investigated a black-box targeted adversarial attack method called VCP-Attack for large-scale vision-language models, combining contrastive learning and PCA subspace projection to generate transferable attack samples.
VCU-Bridge: Hierarchical Visual Connotation Understanding via Semantic Bridging
Ming Zhong (Peng Cheng Laboratory), Wentao Zhang (Zhejiang University)
Explainability and InterpretabilityLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the VCU-Bridge framework, defining a three-layer visual meaning reasoning process from the perception layer (L perc) to the semantic bridge layer (L bridge) and finally to the connotation layer (L conn), and designed the HVCU-Bench benchmark based on this.
VDE: Training-Free Accelerating Rectified Flow Model via Velocity Decomposition and Estimation
Junwen Tan (South China University of Technology), Shuangping Huang (South China University of Technology)
GenerationComputational EfficiencyFlow-based ModelRectified FlowImageVideo
🎯 What it does: This paper proposes a training-agnostic acceleration method called VDE (Velocity Decomposition and Estimation) to speed up the sampling process of Rectified Flow models;
VDFE: Difference-Aware 3D Scene Editing with Non-Intrusive Video Diffusion Priors for Multi-View Consistency and Efficiency
Chao Zhang (Xidian University), Siqi Yu (Xidian University)
GenerationDiffusion modelFlow-based ModelGaussian SplattingVideoPoint Cloud
🎯 What it does: This paper proposes a difference-aware 3D scene editing framework called VDFE based on video diffusion model priors, which utilizes the 3D Gaussian Splatting model for efficient and controllable 3D editing, and achieves multi-view consistency through pseudo video editing.
VDOT: Efficient Unified Video Creation via Optimal Transport Distillation
Yutong Wang (University of Sydney), Xinyuan Chen (Shanghai AI Laboratory)
GenerationKnowledge DistillationGenerative Adversarial NetworkVideo
🎯 What it does: Proposed VDOT, an efficient unified video generation model that leverages Optimal Transport Distillation (OTD);
VecAttention: Vector-wise Sparse Attention for Accelerating Long Context Inference
Anmin Liu (Peking University), Tao Xie (Peking University)
Computational EfficiencyTransformerVideo
🎯 What it does: Propose VecAttention, a vector-level sparse attention framework designed for long video reasoning
VecGlypher: Unified Vector Glyph Generation with Language Models
Xiaoke Huang (Meta AI), Xiao Han (Meta AI)
GenerationTransformerLarge Language ModelMultimodality
🎯 What it does: VecGlypher is a multimodal language model capable of directly generating editable SVG vector glyphs from text descriptions or image references.
Vector Prism: Animating Vector Graphics by Stratifying Semantic Structure
Jooyeol Yun, Jaegul Choo (KAIST)
GenerationTransformerVision Language ModelMultimodality
🎯 What it does: Proposed the Vector Prism framework, which leverages multi-perspective weak labels and statistical inference to recover the semantic structure of SVG, enabling VLM to plan and generate animations at the semantic level.
VectorArk: Learning Practical Image Vectorization with Rounded Polygon Representation
Tarun Gehlaut (Adobe), Vineet Batra (Adobe)
Image TranslationTransformerLarge Language ModelVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Proposed a image vectorization framework VectorArk based on a multimodal large language model (InternVL), which can automatically convert images from the real world or text generation into high-quality, editable SVGs;
Velox: Learning Representations of 4D Geometry and Appearance
Anagh Malik (Apple), Jen-Hao Rick Chang (Apple)
Object TrackingGenerationData SynthesisCompressionRepresentation LearningTransformerDiffusion modelGaussian SplattingVideoPoint Cloud
🎯 What it does: This paper proposes the Velox framework, which learns compressed dynamic tokens for 4D (spatiotemporal) geometry and appearance, and trains two decoders (4D surface flow matching decoder and 3D Gaussian decoder) on this representation, achieving multi-task capabilities including video-to-4D generation, 3D tracking, and cloth simulation.
VEMamba: Efficient Isotropic Reconstruction of Volume Electron Microscopy with Axial-Lateral Consistent Mamba
Longmi Gao (Nanjing University of Aeronautics and Astronautics), Pan Gao (Nanjing University of Aeronautics and Astronautics)
RestorationConvolutional Neural NetworkContrastive LearningBiomedical Data
🎯 What it does: Proposes the VEMamba framework for efficiently restoring isotropic resolution from anisotropic volumetric electron microscopy data.
VENI: Variational Encoder for Natural Illumination
Paul Walker (Friedrich-Alexander-Universitat Erlangen-Nurnberg), Bernhard Egger (University of York)
GenerationRepresentation LearningTransformerAuto EncoderImage
🎯 What it does: Designed and implemented a rotation-equivariant variational autoencoder to learn high dynamic range spherical representations of natural illumination.
Venus: Benchmarking and Empowering Multimodal Large Language Models for Aesthetic Guidance and Cropping
Tianxiang Du (Peking University), Yuxin Peng (Peking University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose a two-phase framework Venus, which first trains an MLLM to achieve aesthetic guidance using the AesGuide dataset, then activates its cropping capability through Chain-of-Thought aesthetic reasoning, enabling interpretable interactive aesthetic cropping.
Verifying Neural Network Robustness with Dual Perturbations
Hai Duong (George Mason University), ThanhVu Nguyen (George Mason University)
ClassificationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Propose the VeriDou framework to verify the robustness of neural networks under dual perturbations, where both convolutional (spatially correlated) perturbations and independent pixel noise coexist.
VerseCrafter: Dynamic Realistic Video World Model with 4D Geometric Control
Sixiao Zheng (Fudan University), Yanwei Fu
SegmentationGenerationDepth EstimationDiffusion modelGaussian SplattingWorld ModelVideoTextPoint Cloud
🎯 What it does: Built a dynamic video world model called VerseCrafter based on 4D geometric control, which can achieve precise control of the camera and multiple objects and generate high-quality videos.
VES-RFT: Rewarding Visual Evidence Sensitivity to Mitigate Hallucinations in Large Vision-Language Models
Xuehe Hou, Shengjin Wang (Tsinghua University)
Explainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningMultimodality
🎯 What it does: Propose a reinforcement learning fine-tuning framework called VES-RFT based on Visual Evidence Sensitivity (VES), which leverages the difference in model uncertainty under image-present and image-absent conditions to reward models for relying more on visual evidence, thereby reducing object hallucinations in large vision-language models.
VesMamba: 3D Pulmonary Vessel Segmentation from CT images via Mamba with Structural Perception and Scale-aware Filtering
Zhipeng Liu (Shenzhen University), Jing Qin (Hong Kong Polytechnic University)
SegmentationBiomedical DataComputed Tomography
🎯 What it does: This paper proposes a 3D lung vessel segmentation model named VesMamba, which can automatically extract vascular structures from CT images;
VGA-Bench: A Unified Benchmark and Multi-Model Framework for Video Aesthetics and Generation Quality Evaluation
Longteng Jiang (Ant Group), Xin Jin (State Key Laboratory of General Artificial Intelligence, BIGAI)
GenerationData SynthesisTransformerPrompt EngineeringVision Language ModelDiffusion modelVideoTextBenchmark
🎯 What it does: Constructed the VGA-Bench unified benchmark covering 52 fine-grained dimensions, designed 1016 prompts and generated over 60,000 videos.
VGA: Empowering Aerial-Ground Localization by Visual Geometry Alignment
Tao Jun Lin (Australian National University), Hongdong Li (Australian National University)
Pose EstimationOptimizationTransformerSimultaneous Localization and MappingImage
🎯 What it does: This paper proposes the VGA framework, which jointly learns camera calibration, gravity direction, and BEV projection to achieve 6-DoF localization for drones and ground cameras.
VGent: Visual Grounding via Modular Design for Disentangling Reasoning and Prediction
Weitai Kang (University of Illinois Chicago), Kangning Liu (Adobe)
Object DetectionSegmentationTransformerLarge Language ModelReinforcement LearningImageTextMultimodality
🎯 What it does: Propose the VGent modular encoder-decoder framework, which uses a frozen multimodal large language model (MLLM) for high-level reasoning, while the decoder leverages candidate boxes generated by detectors for low-level localization, avoiding inefficiencies and hallucinations from autoregressive generation.
VGG-T$^3$: Offline Feed-Forward 3D Reconstruction at Scale
Sven Elflein (NVIDIA), Aljosa Osep (NVIDIA)
Pose EstimationDepth EstimationTransformerImageBenchmark
🎯 What it does: Propose an offline feed-forward 3D reconstruction model named VGG-T3, which compresses variable-length key-value (KV) spaces into fixed-size MLPs via test-time training in global self-attention, achieving linear time complexity in input view numbers, enabling large-scale reconstruction of thousands of unposed tourist images within one minute.
VGGDrive: Empowering Vision-Language Models with Cross-View Geometric Grounding for Autonomous Driving
Jie Wang (Tianjin University), Long Chen (Xiaomi EV)
Autonomous DrivingTransformerVision Language ModelImagePoint Cloud
🎯 What it does: Propose the VGGDrive framework, integrating the visual 3D foundational model VGGT into a vision-language model (VLM) to enhance performance in autonomous driving tasks.
VGGT-360: Geometry-Consistent Zero-Shot Panoramic Depth Estimation
Jiayi Yuan (Singapore University of Technology and Design), Na Zhao (Singapore University of Technology and Design)
Depth EstimationTransformerImage
🎯 What it does: Proposes VGGT-360, a training-free, geometrically consistent omnidirectional depth estimation framework.
VGGT-Det: Mining VGGT Internal Priors for Sensor-Geometry-Free Multi-View Indoor 3D Object Detection
Yang Cao (Hong Kong University of Science and Technology), Dan Xu (Hong Kong University of Science and Technology)
Object DetectionTransformerImage
🎯 What it does: Propose the VGGT-Det framework, leveraging pre-trained semantic and geometric priors from VGGT, achieving precise detection in multi-view indoor 3D object detection tasks without sensor geometry input;
VGGT-ohm
Jianyuan Wang (University of Oxford), Christian Rupprecht (University of Oxford)
Depth EstimationTransformerVideoPoint CloudBenchmark
🎯 What it does: Proposes VGGT, a scalable feed-forward 3D reconstruction model that can efficiently estimate camera parameters and depth maps in both static and dynamic scenes;
VGGT-Segmentor: Geometry-Enhanced Cross-View Segmentation
Yulu Gao (Beihang University), Si Liu (Beihang University)
SegmentationTransformerImageVideo
🎯 What it does: Proposes VGGT-Segmentor (VGGT-S), a framework that integrates geometric information into cross-view instance segmentation, leveraging the geometric consistency features of VGGT with a three-stage joint segmentation head to achieve precise segmentation from egocentric to exocentric (and vice versa).
VIAFormer: Voxel-Image Alignment Transformer for High-Fidelity Voxel Refinement
Tiancheng Fang (Shanghai Jiao Tong University), Fan Wu (Shanghai Jiao Tong University)
RestorationTransformerFlow-based ModelAuto EncoderImageMultimodality
🎯 What it does: Proposes a Voxel-Image Alignment Transformer called VIAFormer for high-fidelity restoration of noisy, missing, or distorted voxel grids under multi-view image guidance.
Vibe Spaces for Creatively Connecting and Expressing Visual Concepts
Huzheng Yang (University of Pennsylvania), Jianbo Shi (University of Pennsylvania)
GenerationGraph Neural NetworkVision Language ModelDiffusion modelContrastive LearningImage
🎯 What it does: Proposes the Vibe Blending task, generating creative image blends along the geometric shortest path in pre-trained feature spaces such as CLIP by learning a low-dimensional graph manifold (Vibe Space).
ViBES: A Conversational Agent with Behaviorally-Intelligent 3D Virtual Body
Juze Zhang (Stanford University), Ehsan Adeli (Stanford University)
GenerationData SynthesisRobotic IntelligenceTransformerLarge Language ModelMixture of ExpertsVision-Language-Action ModelVideoTextMeshAudio
🎯 What it does: Built ViBES, a 3D conversational agent capable of simultaneously understanding speech and text while generating corresponding facial expressions and full-body movements.
VibeToken: Scaling 1D Image Tokenizers and Autoregressive Models for Dynamic Resolution Generations
Maitreya Patel (Arizona State University), Lingjuan Lv (SonyAI)
GenerationComputational EfficiencyTransformerImage
🎯 What it does: Propose a 1D Transformer image tokenizer VibeToken and its corresponding autoregressive generative model VibeToken-Gen, which can adapt to arbitrary resolutions and aspect ratios, achieving efficient multi-resolution image generation.
Video Generation with Stable Transparency via Shiftable RGB-A Distribution Learner
Haotian Dong (Tianjin University), Di Lin (Tianjin University)
GenerationTransformerDiffusion modelAuto EncoderVideo
🎯 What it does: Studied a method capable of generating high-quality videos with transparent channels (RGB-A).
Video Panels for Long Video Understanding
Lars Doorenbos (University of Bonn), Juergen Gall (University of Bonn)
Vision Language ModelVideoBenchmark
🎯 What it does: Propose a no-training, no-parameter, long-video-oriented visual prompting method—concatenating multiple frames into a panel image and inputting it into existing video-language models;
Video-as-Answer: Predict and Generate Next Video Event with Joint-GRPO
Junhao Cheng (City University of Hong Kong), Jing Liao (City University of Hong Kong)
GenerationReinforcement LearningVision Language ModelDiffusion modelVideoMultimodality
🎯 What it does: Proposes the Video Next Event Prediction (VNEP) task, answering questions using video as answers rather than text;
Video-CoE: Reinforcing Video Event Prediction via Chain of Events
Qile Su (AMAP, Alibaba Group), Xiangxiang Chu (AMAP, Alibaba Group)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextChain-of-Thought
🎯 What it does: Proposed and implemented the Video-CoE framework, which leverages event chains to enhance the logical reasoning and visual utilization of multimodal large language models (MLLM) for video event prediction (VEP).
Video-Only ToM: Enhancing Theory of Mind in Multimodal Large Language Models
Siqi Liu (University of Science and Technology Beijing), Jiansheng Chen (University of Science and Technology Beijing)
Explainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodality
🎯 What it does: This paper proposes VisionToM, a framework based on visual intervention, which utilizes learnable intervention vectors to align visual representations with correct semantic targets, thereby regulating attention within multimodal large language models (MLLMs) and enhancing the model's Theory of Mind (ToM) reasoning capabilities in video-only scenarios.
Video2Robo: 3DGS-based Synthetic Data from One Video Enables Scalable Robot Learning
Yinan Deng (Beijing Institute of Technology), Yufeng Yue (Beijing Institute of Technology)
Data SynthesisPose EstimationDepth EstimationRobotic IntelligenceDiffusion modelGaussian SplattingImageVideo
🎯 What it does: Achieve 3D reconstruction, 6D pose tracking, and synthesize diverse, realistic robot learning data by utilizing 3D Gaussian Splatting (3DGS) through a single RGB human demonstration video; subsequently, train visual-motor policies using these synthetic data and deploy them directly on real robots.
VideoARM: Agentic Reasoning over Hierarchical Memory for Long-Form Video Understanding
Yufei Yin (Hangzhou Dianzi University), Zhou Yu (Hangzhou Dianzi University)
TransformerLarge Language ModelAgentic AIVision-Language-Action ModelVideoMultimodalityBenchmark
🎯 What it does: This paper proposes a long video understanding framework based on Agentic Reasoning over Hierarchical Memory (VideoARM), adopting an Observe-Think-Act-Memorize cycle to dynamically construct multi-modal memory and adaptively invoke multi-modal tools on this basis, achieving content understanding from coarse to fine granularity.
VideoAuto-R1: Video Auto Reasoning via Thinking Once, Answering Twice
Shuming Liu (Meta AI), Yunyang Xiong (Meta AI)
Large Language ModelReinforcement LearningVision Language ModelVideoMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the VideoAuto-R1 framework, adopting a 'think once, answer twice' training mode and adaptive early exit mechanism to achieve automatic reasoning for video understanding tasks.
VideoChat-M1: Collaborative Policy Planning for Video Understanding via Multi-Agent Reinforcement Learning
Boyu Chen (Chinese Academy of Sciences), Yali Wang (Chinese Academy of Sciences)
RecognitionTransformerLarge Language ModelReinforcement LearningAgentic AIVideoText
🎯 What it does: Propose a multi-agent collaborative video understanding framework named VideoChat‑M1, which achieves dynamic tool calling and strategy optimization through collaborative strategy planning (CPP) and multi-agent reinforcement learning (MARL).
VideoCoF: Unified Video Editing with Temporal Reasoner
Xiangpeng Yang (University of Technology Sydney), Qiang Wu (University of Technology Sydney)
GenerationTransformerLarge Language ModelDiffusion modelAuto EncoderVideoMultimodalityBenchmarkChain-of-ThoughtOrdinary Differential Equation
🎯 What it does: Propose the VideoCoF framework, leveraging Chain-of-Frames (insight → reasoning → editing) to achieve unified and precise video editing;
VideoFusion: A Spatio-Temporal Collaborative Network for Multi-modal Video Fusion
Linfeng Tang (Wuhan University), Jiayi Ma (Wuhan University)
Data SynthesisRepresentation LearningConvolutional Neural NetworkTransformerVideoMultimodality
🎯 What it does: Constructed a multimodal multiscenario video dataset M3SVD (220 synchronized thermal infrared-visible light videos, totaling 153,797 frames), and proposed the VideoFusion network to achieve multimodal video fusion.
VideoITG: Multimodal Video Understanding with Instructed Temporal Grounding
Shihao Wang (Hong Kong Polytechnic University), Zhiding Yu (NVIDIA)
Computational EfficiencyTransformerLarge Language ModelPrompt EngineeringVideoTextMultimodality
🎯 What it does: Designed and implemented the VideoITG framework, which improves the efficiency of multi-modal video understanding through instruction-driven temporal frame selection, and constructed the VideoITG-40K dataset via the VidThinker automated annotation pipeline.
VideoMaMa: Mask-Guided Video Matting via Generative Prior
Sangbeom Lim (Korea University), Joon-Young Lee (Adobe Research)
SegmentationGenerationData SynthesisDiffusion modelAuto EncoderVideoBenchmark
🎯 What it does: Propose VideoMaMa, a method based on diffusion models for converting binary segmentation masks to continuous alpha matting, and use it to generate pseudo-labels to construct a large-scale MA-V video matting dataset.
VidEoMT: Your ViT is Secretly Also a Video Segmentation Model
Narges Norouzi (Eindhoven University of Technology), Daan de Geus (Eindhoven University of Technology)
SegmentationTransformerVideo
🎯 What it does: Proposed a pure encoder (encoder-only) video segmentation framework called VidEoMT, which can achieve segmentation and temporal association without using any dedicated tracking modules.
VideoNet: A Large-Scale Dataset for Domain-Specific Action Recognition
Tanush Yadav (University of Washington), Ranjay Krishna (University of Washington)
RecognitionData SynthesisLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes VideoNet, a domain-specific action recognition benchmark covering 38 domains and 1,087 fine-grained actions, and constructs a large-scale training dataset (approximately 160k–500k video question-answer pairs). Through automated web crawling, video localization (Gemini 2.5 Flash), transcription (WhisperX), and multi-stage human verification, video clips, negative sample generation, and action definition writing were completed. Subsequently, the zero-shot and few-shot performance of various audio-visual models (Molmo2, InternVL3, LLaVA, Qwen2.5VL, Gemini, GPT‑4, GPT‑5, etc.) on this benchmark was evaluated, and Molmo2‑4B was fine-tuned on the training set, significantly enhancing domain-specific action recognition capabilities.
VideoRealBench: A Chain-of-Thought Realism Evaluation Benchmark for Generated Human-Centric Videos
Min Yang (Nanjing University), Limin Wang (Nanjing University)
Explainability and InterpretabilitySupervised Fine-TuningVision Language ModelVideoTextBenchmarkChain-of-Thought
🎯 What it does: Constructed a high-quality human-centered video authenticity evaluation benchmark, VideoRealBench, which includes the VideoRealDataset dataset and an automatic evaluator, VideoRealEval.
VideoSeek: Long-Horizon Video Agent with Tool-Guided Seeking
Jingyang Lin (AMD), Emad Barsoum (AMD)
RetrievalExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelAgentic AIVision-Language-Action ModelVideoTextChain-of-Thought
🎯 What it does: Proposes VideoSeek, a long-sequence video agent that actively identifies key information relevant to answering questions through video logic flow.
VideoSSR: Video Self-Supervised Reinforcement Learning
Zefeng He (Shanghai Artificial Intelligence Laboratory), Yu Cheng (Chinese University of Hong Kong)
TransformerLarge Language ModelReinforcement LearningVideoMultimodalityBenchmark
🎯 What it does: Proposed a self-supervised reinforcement learning framework called VideoSSR that leverages video's own information, addressing the data scarcity issue in training large multimodal language models for video understanding.
VideoWeaver: Multimodal Multi-View Video-to-Video Transfer for Embodied Agents
George Eskandar, Ziyuan Liu (Huawei Heisenberg Research Center)
GenerationData SynthesisRobotic IntelligenceTransformerMixture of ExpertsFlow-based ModelRectified FlowAuto EncoderVideoTextMultimodalityPoint Cloud
🎯 What it does: Propose VideoWeaver, a flow-based multi-modal multi-view video-to-video (V2V) conversion framework that can convert robot demonstration videos captured by multiple synchronized cameras into arbitrary styles specified by text prompts, while maintaining cross-view consistency.
VideoWorld 2: Learning Transferable Knowledge from Real-world Videos
Zhongwei Ren (Beijing Jiaotong University), Xiaojie Jin (ByteDance Seed)
Domain AdaptationRepresentation LearningRobotic IntelligenceTransformerDiffusion modelAuto EncoderVideo
🎯 What it does: This paper proposes VideoWorld 2, which can learn transferable task knowledge from unannotated real-world videos and execute complex long-horizon tasks in new environments.
VidPrism: Heterogeneous Mixture of Experts for Image-to-Video Transfer
Rui Lin (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)
RecognitionTransformerMixture of ExpertsVision Language ModelVideo
🎯 What it does: Propose VidPrism, a heterogeneous Mixture-of-Experts (MoE) framework, to migrate pre-trained vision-language models to video understanding tasks, addressing the homogenization problem of traditional MoE experts.
VidTAG: Temporally Aligned Video to GPS Geolocalization with Denoising Sequence Prediction at a Global Scale
Parth Parag Kulkarni (University of Central Florida), Mubarak Shah (University of Central Florida)
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningVideoMultimodalitySequential
🎯 What it does: Propose a global video localization framework called VidTAG, which generates fine-grained trajectories using frame-level GPS retrieval.
View-Aware Semantic Alignment for Aerial-Ground Person Re-Identification
Quan Zhang (Sun Yat-sen University), Jianhuang Lai (Sun Yat-sen University)
RecognitionRetrievalGraph Neural NetworkTransformerMixture of ExpertsImage
🎯 What it does: Proposes a perspective-aware semantic alignment framework named ViSA for aerial-ground person re-identification. The framework achieves separation and fusion of viewpoint-invariant and viewpoint-specific features through a perspective-decoupled Transformer, an Expert-Driven Token Generation Module (ETGM), and a Dual-Branch Local Fusion Module (DLFM).
ViHOI: Human-Object Interaction Synthesis with Visual Priors
Songjin Cai (South China University of Technology), Changxing Ding (South China University of Technology)
GenerationData SynthesisVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Design the ViHOI framework, which leverages a Vision-Language Model (VLM) to extract visual and textual priors, compresses them into compact tokens via a Q-Former adapter, and integrates with a diffusion generation model. During training, real-action rendered images serve as visual priors, while during inference, reference images are synthesized using text-to-image models, enabling the generation of more realistic and physically feasible 3D human-object interaction actions.
ViKey: Enhancing Temporal Understanding in Videos via Visual Prompting
Yeonkyung Lee (Yonsei University), Seong Jae Hwang (Yonsei University)
TransformerPrompt EngineeringVision Language ModelVideo
🎯 What it does: Propose the VIKEY framework, which enhances the temporal reasoning capabilities of existing VideoLLMs without additional training by overlaying explicit frame numbering (e.g., 'frame #01') as visual prompts on each video frame, combined with keyword-frame mapping (KFM) technology.
ViLearn: Accelerating Training Convergence of Image-to-3D Generation via Visibility Learning
Rui Chen, Ping Tan
GenerationTransformerDiffusion modelAuto EncoderImageMesh
🎯 What it does: Propose the ViLearn method, which significantly accelerates the training convergence of single-image-to-3D generation through Visibility Learning.
ViLoMem: Agentic Learner with Grow-and-Refine Multimodal Semantic Memory
Weihao Bo (Nanjing University of Science and Technology), Zechao Li (Nanjing University of Science and Technology)
Representation LearningTransformerLarge Language ModelVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposed the ViLoMem dual-stream memory framework, which utilizes visual and logical streams to separately record and retrieve error patterns, supporting progressive learning in multi-modal LLMs.
VIMCAN: Visual-Inertial 3D Human Pose Estimation with Hybrid Mamba-Cross-Attention Network
Zepeng Yang (Beihang University), Bin Li (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
Pose EstimationMultimodality
🎯 What it does: Proposed a visual-inertial hybrid network named VIMCAN for 3D human pose estimation.
Vinedresser3D: Towards Agentic Text-guided 3D Editing
Yankuan Chi (Hong Kong University of Science and Technology), James Matthew Rehg (University of Illinois Urbana Champaign)
GenerationTransformerLarge Language ModelAgentic AIPrompt EngineeringFlow-based ModelRectified FlowImageTextMultimodalityMesh
🎯 What it does: Proposes Vinedresser3D, a text-guided 3D editing framework based on agents, which utilizes a multi-modal large language model (Gemini-2.5-flash) to generate text and image instructions, automatically detects editing regions, and combines 3D segmentation (PartField), image editing models (Nano Banana), and local 3D generation models (Trellis) for 3D spatial editing based on inverse flow;
VinQA: Visual Elements Interleaved Long-form Answer Generation for Real-World Multimodal Document QA
Young Rok Jang (LG AI Research), Stanley Jungkyu Choi (LG AI Research)
GenerationTransformerLarge Language ModelSupervised Fine-TuningMultimodalityBenchmark
🎯 What it does: Constructed the VinQA dataset and proposed two methods to convert original page images into inputs for multi-modal large language models (Page Encoding and Modality Encoding). Designed two evaluation frameworks, M-GroSE and Visual G-Eval, to complete the task of explicitly embedding visual elements into long-form answer generation.
VINS-120K: Ultra High-Resolution Image Editing with A Large-Scale Dataset
Zhizhou Chen (Nanjing University), Ying Tai (Nanjing University)
GenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelImageBenchmark
🎯 What it does: This paper proposes VINS-120K, a large-scale dataset containing 120,000 entries of 4K+ ultra-high-resolution instruction editing data, and develops a high-frequency perception post-adaptation strategy based on this dataset, significantly improving the detail fidelity and instruction following capability in ultra-high-resolution image editing.
VIRAL: Visual Sim-to-Real at Scale for Humanoid Loco-Manipulation
Tairan He (NVIDIA), Yuke Zhu (NVIDIA)
Domain AdaptationRobotic IntelligenceTransformerReinforcement LearningImageTabular
🎯 What it does: On a Unitree G1 humanoid robot, a fully RGB-image-based visual policy was trained using a teacher-student structure, enabling the robot to autonomously complete multi-stage long-sequence 'locomo-manipulation' tasks (e.g., walking, grasping, placing) and execute them directly in real-world environments without any tuning.
ViRC: Enhancing Visual Interleaved Mathematical CoT with Reason Chunking
Lihong Wang (Jilin University), Zhe Li (Ant Digital Technologies, Ant Group)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the VIRC framework, which decomposes multimodal mathematical reasoning into key reasoning units (CRUs) using the 'Reason Chunking' mechanism, and constructed the CRUX dataset along with a three-stage training strategy;
VIRD: View-Invariant Representation through Dual-Axis Transformation for Cross-View Pose Estimation
Juhye Park (KAIST), Hyun Myung (KAIST)
Pose EstimationAutonomous DrivingConvolutional Neural NetworkImage
🎯 What it does: Proposes a cross-perspective pose estimation method based on dual-axis transformation (VIRD), achieving precise alignment and localization between ground camera images and satellite images without relying on directional priors.