ICCV 2025 Papers — Page 25
IEEE/CVF International Conference on Computer Vision · 2701 papers
Towards Human-like Virtual Beings: Simulating Human Behavior in 3D Scenes
Chen Liang (Zhejiang University), Yi Yang (Zhejiang University)
Robotic IntelligenceTransformerLarge Language ModelAgentic AI
🎯 What it does: Designed and implemented an autonomous intelligent agent ACTOR based on large language models (LLM), capable of executing high-level, long-term abstract goal-driven behaviors in real 3D home environments, and achieving dynamic environment adaptation through a perception-planning-execution loop. A large-scale scene perception and action-rich BEHAVIORHUB dataset was also constructed for training and evaluating such behavior simulations.
Towards Immersive Human-X Interaction: A Real-Time Framework for Physically Plausible Motion Synthesis
Kaiyang Ji (ShanghaiTech University), Jingya Wang (ShanghaiTech University)
Robotic IntelligenceReinforcement LearningDiffusion modelMultimodality
🎯 What it does: The Human-X framework is proposed, achieving real-time, physically feasible interactive motion synthesis between humans, virtual characters, and robots in VR/AR environments.
Towards Long-Horizon Vision-Language-Action System: Reasoning, Acting and Memory
Daixun Li (Xidian University), Yunsong Li (Xidian University)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningMixture of ExpertsVision-Language-Action ModelDiffusion modelMultimodalityChain-of-Thought
🎯 What it does: The paper presents a hierarchical entity intelligence system named MindExplore, which combines reasoning, execution, and memory modules to accomplish long-term tasks in highly dynamic environments such as deserts.
Towards More Diverse and Challenging Pre-training for Point Cloud Learning: Self-Supervised Cross Reconstruction with Decoupled Views
Xiangdong Zhang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
SegmentationRepresentation LearningTransformerPoint Cloud
🎯 What it does: Research on self-supervised pre-training of 3D point clouds, proposing a cross-reconstruction framework called Point-PQAE.
Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation
Kaining Ying (Fudan University), Yu-Gang Jiang (Fudan University)
SegmentationTransformerLarge Language ModelSupervised Fine-TuningVideoTextMultimodalityAudio
🎯 What it does: An OmniAVS dataset covering 2,104 videos and 61,095 multimodal descriptions is proposed, along with an end-to-end speech-visual segmentation baseline designed based on a multimodal large language model (OISA).
Towards Open-World Generation of Stereo Images and Unsupervised Matching
Feng Qiao (Washington University in St. Louis), Nathan Jacobs (Washington University in St. Louis)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: The GenStereo framework is proposed, utilizing diffusion models to achieve high-quality stereo image generation in the open-source world, and further applied to unsupervised stereo matching training.
Towards Performance Consistency in Multi-Level Model Collaboration
Qi Li (National University of Singapore), Xinchao Wang (National University of Singapore)
ClassificationConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: This study investigates the performance consistency between parameter-level model merging and prediction-level model ensemble, and proposes the NeuLig verification framework to achieve performance alignment under multi-model collaboration.
Towards Privacy-preserved Pre-training of Remote Sensing Foundation Models with Federated Mutual-guidance Learning
Jieyi Tan (Wuhan University), Yansheng Li (Wuhan University)
Object DetectionSegmentationFederated LearningSafty and PrivacyKnowledge DistillationContrastive LearningImage
🎯 What it does: The FedSense framework is proposed to enable multiple institutions to collaboratively pre-train remote sensing foundational models without sharing raw remote sensing data.
Towards Real Unsupervised Anomaly Detection Via Confident Meta-Learning
Muhammad Aqeel (University of Verona), Francesco Setti (Qualyco S.r.l.)
Anomaly DetectionMeta LearningImage
🎯 What it does: A new training framework called CoMet is proposed, which allows anomaly detection models to be trained on real unsupervised datasets containing unlabeled anomalous samples.
Towards Robust Defense against Customization via Protective Perturbation Resistant to Diffusion-based Purification
Wenkui Yang (Chinese Academy of Sciences), Ran He (Chinese Academy of Sciences)
GenerationData SynthesisAdversarial AttackDiffusion modelImage
🎯 What it does: A protection perturbation method called AntiPure is proposed to resist purification in the diffusion model purification-customization (P-C) process, which can significantly interfere with the quality of subsequent customized generation while maintaining low perceptual differences.
Towards Robustness of Person Search against Corruptions
Woojung Son (Korea Advanced Institute of Science and Technology), Sung-Eui Yoon (Korea Advanced Institute of Science and Technology)
RecognitionObject DetectionRetrievalConvolutional Neural NetworkTransformerSupervised Fine-TuningImageBenchmark
🎯 What it does: A robust evaluation framework for person search tasks under image corruption conditions is proposed, along with two benchmark datasets CUHK‑SYSU‑C and PRW‑C. Two techniques, foreground-aware enhancement and robust proposal regularization, are designed to improve model robustness.
Towards Safer and Understandable Driver Intention Prediction
Mukilan Karuppasamy (Indian Institute of Information Technology Hyderabad), C V Jawahar (Politecnico di Torino)
Autonomous DrivingExplainability and InterpretabilityTransformerVideoMultimodality
🎯 What it does: An interpretable driving intention prediction model VCBM is constructed, and a multimodal explanation dataset DAAD-X is proposed.
Towards Scalable Spatial Intelligence via 2D-to-3D Data Lifting
Xingyu Miao (Durham University), Gongjie Zhang (Alibaba Group)
SegmentationData SynthesisDepth EstimationMixture of ExpertsImagePoint Cloud
🎯 What it does: A scalable 2D to 3D data enhancement pipeline is proposed, which converts single-view images into three-dimensional scenes with real scale and realistic appearance through depth estimation, camera calibration, and scale calibration, including point clouds, depth maps, camera poses, and pseudo RGB-D, etc.;
Towards Stabilized and Efficient Diffusion Transformers through Long-Skip-Connections with Spectral Constraints
Guanjie Chen (Shanghai Jiao Tong University), Yu Cheng (Chinese University of Hong Kong)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelImageVideoText
🎯 What it does: This paper proposes the Skip-DiT architecture, which incorporates long skip connections into the Diffusion Transformer (DiT) to enhance feature stability and achieve efficient caching acceleration.
Towards Video Thinking Test: A Holistic Benchmark for Advanced Video Reasoning and Understanding
Yuanhan Zhang (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
TransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmarkChain-of-ThoughtAudio
🎯 What it does: Proposes the Video Thinking Test (Video-TT) benchmark to evaluate the correctness and robustness of video large language models in complex visual storytelling.
Towards Visual Localization Interoperability: Cross-Feature for Collaborative Visual Localization and Mapping
Alberto Jaenal (Ericsson Research), André Mateus (Ericsson Research)
Object DetectionPose EstimationSimultaneous Localization and MappingImage
🎯 What it does: The Cross-Feature method is proposed to establish correspondences between different local feature algorithms, achieving interoperability in visual localization.
TPG-INR: Target Prior-Guided Implicit 3D CT Reconstruction for Enhanced Sparse-view Imaging
Qinglei Cao (Southwest University), Xiaoqin Tang (Southwest University)
RestorationOptimizationNeural Radiance FieldImageComputed Tomography
🎯 What it does: A target prior-guided implicit 3D CT reconstruction framework TPG-INR is proposed, which quickly estimates the target prior using sparse angle projections to guide voxel sampling and structural encoding, achieving high-quality reconstruction.
TR-PTS: Task-Relevant Parameter and Token Selection for Efficient Tuning
Siqi Luo (Shanghai Jiao Tong University), Xiaohong Liu (Shanghai Jiao Tong University)
ClassificationComputational EfficiencyTransformerSupervised Fine-TuningImage
🎯 What it does: A TR-PTS framework is proposed, achieving efficient fine-tuning in the ViT model through task-related parameter selection and token selection.
TRACE: Learning 3D Gaussian Physical Dynamics from Multi-view Videos
Jinxi Li (Hong Kong Polytechnic University), Bo Yang (Hong Kong Polytechnic University)
Object DetectionSegmentationGenerationGaussian SplattingVideoPhysics RelatedOrdinary Differential Equation
🎯 What it does: This work proposes the TRACE framework, which utilizes multi-view RGB videos for unsupervised learning of 3D Gaussian particles as rigid body particles, directly predicting their translational-rotational dynamic parameters, thereby achieving future frame extrapolation.
Trace3D: Consistent Segmentation Lifting via Gaussian Instance Tracing
Hongyu Shen (Beijing Institute of Technology), Siyuan Huang (State Key Laboratory of General Artificial Intelligence, BIGAI)
Object DetectionSegmentationContrastive LearningGaussian SplattingImage
🎯 What it does: This paper proposes a method called Trace3D that elevates 2D segmentation to high-precision 3D Gaussian Splatting, capable of achieving consistent 3D semantic/instance segmentation from single or multiple viewpoints, and enabling object extraction and scene editing.
Tracing Copied Pixels and Regularizing Patch Affinity in Copy Detection
Yichen Lu (Ant Group), Peng Zhang (Ant Group)
RecognitionImage TranslationRetrievalOptimizationExplainability and InterpretabilityTransformerContrastive LearningImage
🎯 What it does: A pixel coordinate tracking module (PixTrace) utilizing pixel coordinate tracking and a geometry-guided contrastive loss (CopyNCE) has been designed and implemented to enhance fine-grained alignment and matching of edited regions in image copy detection (ICD).
TrackAny3D: Transferring Pretrained 3D Models for Category-unified 3D Point Cloud Tracking
Mengmeng Wang (Zhejiang University of Technology), Feng Xia (RMIT University)
Object TrackingAutonomous DrivingTransformerMixture of ExpertsPoint Cloud
🎯 What it does: This paper proposes the TrackAny3D framework, which implements the transfer of large-scale pre-trained 3D point cloud models to a category-unified single-object tracking task.
Tracking Tiny Drones against Clutter: Large-Scale Infrared Benchmark with Motion-Centric Adaptive Algorithm
Jiahao Zhang (Beijing University of Technology), Gang Wang (Beijing University of Technology)
Object TrackingConvolutional Neural NetworkVideoBenchmark
🎯 What it does: A large-scale infrared small drone tracking dataset TDTIV has been constructed, and a Motion-Centric Adaptive Tracker (MCATrack) has been proposed for tracking very small targets.
TrackVerse: A Large-Scale Object-Centric Video Dataset for Image-Level Representation Learning
Yibing Wei (University of Wisconsin Madison), Pedro Morgado (University of Wisconsin Madison)
Object DetectionObject TrackingRepresentation LearningTransformerContrastive LearningVideo
🎯 What it does: This paper constructs a large-scale, object-centered video dataset called TrackVerse and proposes a mutation-aware contrastive learning framework based on data augmentation parameters to achieve unsupervised representation learning of object state changes.
Trade-offs in Image Generation: How Do Different Dimensions Interact?
Sicheng Zhang (Khalifa University), Zhichao Lu (City University of Hong Kong)
GenerationTransformerVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes TRIG-Bench, a multi-dimensional evaluation benchmark covering three major tasks: text-to-image, image editing, and topic generation, with 10 dimensions and 132 dimension pairs.
TrafficLoc: Localizing Traffic Surveillance Cameras in 3D Scenes
Yan Xia (University of Science and Technology of China), Daniel Cremers (Technical University of Munich)
Pose EstimationAutonomous DrivingTransformerContrastive LearningImagePoint Cloud
🎯 What it does: This study investigates the problem of locating traffic monitoring cameras in a three-dimensional reference point cloud and proposes a coarse-to-fine image-point cloud registration framework called TrafficLoc.
Training-free and Adaptive Sparse Attention for Efficient Long Video Generation
Yifei Xia (Peking University), Bin Cui (Peking University)
GenerationData SynthesisComputational EfficiencyTransformerVideoBenchmark
🎯 What it does: Proposes AdaSpa, a training-independent adaptive sparse attention mechanism for accelerating long video generation.
Training-Free Class Purification for Open-Vocabulary Semantic Segmentation
Qi Chen (Sun Yat-sen University), Xiaohua Xie (Sun Yat-sen University)
SegmentationVision Language ModelImage
🎯 What it does: A training-free category purification framework called FreeCP is proposed to enhance the accuracy of open vocabulary semantic segmentation.
Training-free Generation of Temporally Consistent Rewards from VLMs
Yinuo Zhao (Beijing Institute of Technology), Jian Tang (Beijing Innovation Center of Humanoid Robotics)
Robotic IntelligenceTransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: A training-free, temporally consistent reward generation framework T2-VLM is proposed, which utilizes VLM to automatically decompose sub-goals and tracks the completion status of sub-goals throughout the task process using a particle filter, thereby providing accurate and real-time reward signals for RL training.
Training-free Geometric Image Editing on Diffusion Models
Hanshen Zhu (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
Image TranslationRestorationGenerationDiffusion modelImageBenchmark
🎯 What it does: A training-free diffusion model is proposed for geometric image editing, which separates object transformation, source area restoration, and target area refinement.
Training-Free Industrial Defect Generation with Diffusion Models
Ruyi Xu (National Taiwan University), Wen-Huang Cheng (National Taiwan University)
GenerationData SynthesisAnomaly DetectionDiffusion modelImage
🎯 What it does: A TF-IDG framework is proposed without the need for any model training, achieving diverse, fine-grained, and structurally consistent industrial defect synthesis through image alignment, mask guidance, and texture preservation.
Training-Free Personalization via Retrieval and Reasoning on Fingerprints
Deepayan Das (University of Trento), Elisa Ricci (Fondazione Bruno Kessler)
RecognitionRetrievalTransformerVision Language ModelMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A training-free VLM personalization method R2P is proposed, which identifies user-specific concepts through retrieval and reasoning.
Training-Free Text-Guided Image Editing with Visual Autoregressive Model
Yufei Wang (Snap Research), Jian Wang (Nanyang Technological University)
Image TranslationGenerationVision Language ModelImageBenchmark
🎯 What it does: This paper proposes a training-free text-guided image editing framework called AREdit, based on a Visual Autoregressive Model (VAR), which achieves high-fidelity local edits without explicit inversion while keeping other areas of the original image unchanged.
TrajectoryCrafter: Redirecting Camera Trajectory for Monocular Videos via Diffusion Models
Mark Yu (Tencent), Ying Shan (Tencent)
GenerationData SynthesisDiffusion modelVideoPoint Cloud
🎯 What it does: This paper presents TrajectoryCrafter, a method capable of redirecting camera trajectories for monocular videos and generating high-quality 4D videos.
Trans-Adapter: A Plug-and-Play Framework for Transparent Image Inpainting
Yuekun Dai (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
RestorationDiffusion modelImageBenchmark
🎯 What it does: This paper proposes Trans-Adapter, a pluggable adapter that enables image inpainting based on diffusion models to directly handle RGBA transparent images.
Transformed Low-rank Adaptation via Tensor Decomposition and Its Applications to Text-to-image Models
Zerui Tao (RIKEN Artificial Intelligence Project), Yuki Mitsufuji (Sony Group Corporation)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: A new parameter-efficient fine-tuning method called TLoRA is proposed, which achieves efficient fine-tuning of text-to-image models by applying learnable full-rank transformations on pre-trained weights and then using low-rank tensor decomposition on the residuals.
Transformer-based Tooth Alignment Prediction with Occlusion and Collision Constraints
Zhenxing Dong (Zhejiang University of Technology), Jiazhou Chen (Zhejiang University of Technology)
TransformerPoint Cloud
🎯 What it does: A tooth alignment prediction network based on Swin-Transformer is proposed, which reorganizes 3D point clouds into ordered multi-channel textures and designs new loss functions for occlusal overlap and distance uniformity, achieving fully automated orthodontic tooth alignment prediction.
TransiT: Transient Transformer for Non-line-of-sight Videography
Ruiqian Li (ShanghaiTech University), Jingyi Yu (ShanghaiTech University)
RestorationGenerationData SynthesisTransformerVideo
🎯 What it does: This paper proposes TransiT, a Transformer-based model capable of real-time reconstruction of 64×64 resolution, 10 FPS NLOS video from transient spectra collected by sparse single-pixel SPAD sensors affected by fast scanning.
Translation of Text Embedding via Delta Vector to Suppress Strongly Entangled Content in Text-to-Image Diffusion Models
Eunseo Koh (Sungkyunkwan University), Jae-Pil Heo (Sungkyunkwan University)
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelImageTextBenchmark
🎯 What it does: This paper proposes a method to directly suppress negative content that is highly entangled with keywords by introducing a 'delta vector' in the text embedding space, achieving precise content suppression on Stable Diffusion and personalized T2I models. It further introduces the Selective Suppression with Delta Vector (SSDV) mechanism.
Transparent Vision: A Theory of Hierarchical Invariant Representations
Shuren Qi (Chinese University of Hong Kong), Fenglei Fan (City University of Hong Kong)
ClassificationRepresentation LearningNeural Architecture SearchConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a Hierarchical Invariant Representation (HIR), which achieves controllable invariance to geometric transformations through an interpretable convolution-nonlinearity-pooling-global invariance layer.
TRCE: Towards Reliable Malicious Concept Erasure in Text-to-Image Diffusion Models
Ruidong Chen (Tianjin University), An-An Liu (Tianjin University)
GenerationData SynthesisTransformerDiffusion modelContrastive LearningImageText
🎯 What it does: TRCE proposes a two-stage concept elimination framework to reliably remove malicious concepts in text-to-image diffusion models.
TREAD: Token Routing for Efficient Architecture-agnostic Diffusion Training
Felix Krause (Computer Vision Foundation), Björn Ommer (Computer Vision Foundation)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: This paper proposes a training strategy called TREAD, which reduces computational load and improves generation quality by routing a portion of tokens from early layers to deeper layers during training, while keeping the model architecture unchanged.
Tree Skeletonization from 3D Point Clouds by Denoising Diffusion
Elias Ariel Marks (University of Bonn), Cyrill Stachniss (University of Bonn)
SegmentationGenerationDiffusion modelPoint CloudAgriculture Related
🎯 What it does: This paper proposes a trunk skeleton reconstruction method based on a denoising diffusion probabilistic model, which can predict the positions and flow directions of branch nodes from partially occluded colored point clouds, and constructs a complete trunk skeleton using a minimum spanning tree algorithm.
Tree-NeRV: Efficient Non-Uniform Sampling for Neural Video Representation via Tree-Structured Feature Grids
Jiancheng Zhao (University of Tokyo), Yinqiang Zheng (University of Tokyo)
CompressionRepresentation LearningVideo
🎯 What it does: Proposes Tree-NeRV, a non-uniform sampling neural video representation framework based on binary search trees.
Triad: Empowering LMM-based Anomaly Detection with Expert-guided Region-of-Interest Tokenizer and Manufacturing Process
Yuanze Li (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
Anomaly DetectionTransformerLarge Language ModelVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: This paper proposes an industrial defect detection framework called Triad, based on a large multimodal model, which combines expert-guided ROI tokenization and manufacturing process reasoning to achieve high-precision defect determination in zero/one-shot scenarios.
Trial-Oriented Visual Rearrangement
Yuyi Liu (Institute of Computing Technology Chinese Academy of Sciences), Shuqiang Jiang (Institute of Computing Technology Chinese Academy of Sciences)
Robotic IntelligencePoint Cloud
🎯 What it does: A Trial-Oriented Visual Rearrangement (TOR) framework is proposed, which achieves visual room rearrangement through differential point clouds, interactive evaluation, and experimental modules.
TriDi: Trilateral Diffusion of 3D Humans, Objects, and Interactions
Ilya A. Petrov (University of Tübingen), Gerard Pons-Moll (Max Planck Institute for Informatics)
GenerationData SynthesisPose EstimationTransformerDiffusion modelMultimodality
🎯 What it does: We propose TriDi, a joint probability model based on triadic diffusion that can simultaneously generate human poses, object poses, and person-object interactions, supporting seven conditional sampling modes.
TRKT: Weakly Supervised Dynamic Scene Graph Generation with Temporal-enhanced Relation-aware Knowledge Transferring
Zhu Xu (Peking University), Yang Liu (Peking University)
Object DetectionGenerationTransformerOptical FlowVideo
🎯 What it does: This paper proposes a weakly supervised dynamic scene graph generation method called TRKT, which utilizes single-frame unlocalized scene graphs for supervision. It enhances external detection through relationship awareness and temporal knowledge transfer, thereby generating more accurate pseudo-labels and improving the quality of dynamic scene graphs.
TRNAS: A Training-Free Robust Neural Architecture Search
Yeming Yang (Shenzhen University), Jianqiang Li (Shenzhen University)
OptimizationAdversarial AttackNeural Architecture SearchImageBenchmark
🎯 What it does: This paper proposes a training-free robust neural architecture search method called TRNAS, which quickly searches for robust networks in the DARTS search space using the R-Score evaluation metric and a multi-objective selection strategy.
Trokens: Semantic-Aware Relational Trajectory Tokens for Few-Shot Action Recognition
Pulkit Kumar (University of Maryland), Abhinav Shrivastava (University of Maryland)
RecognitionTransformerContrastive LearningVideo
🎯 What it does: The Trokens method is proposed, which achieves few-shot action recognition through semantically aware trajectory point sampling and associated motion modeling.
Trust but Verify: Programmatic VLM Evaluation in the Wild
Viraj Prabhu (Salesforce AI Research), Ran Xu (OpenAI)
TransformerLarge Language ModelVision Language ModelImageTextBenchmark
🎯 What it does: A verifiable open visual question answering benchmark PROVE has been constructed, using high-fidelity scene graphs and programmatic verification.
TrustMark: Robust Watermarking and Watermark Removal for Arbitrary Resolution Images
Tu Bui (University of Surrey), John Collomosse (Adobe Research)
Image TranslationRestorationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: Proposes two end-to-end deep learning models, TrustMark and ReMark, to achieve invisible watermark embedding and removal for images of any resolution, while supporting controllable robustness.
TruthPrInt: Mitigating Large Vision-Language Models Object Hallucination Via Latent Truthful-Guided Pre-Intervention
Jinhao Duan (Drexel University), Kaidi Xu (Drexel University)
Object DetectionGenerationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes an object hallucination detection and intervention framework called TruthPrInt, which first learns the 'sincere direction' in a latent subspace and then performs preemptive interventions guided by sincerity during inference to reduce object hallucinations in image descriptions.
TryOn-Refiner: Conditional Rectified-flow-based TryOn Refiner for More Accurate Detail Reconstruction
Wen Qian (Alibaba DAMO Academy)
Image TranslationRestorationDiffusion modelFlow-based ModelRectified FlowAuto EncoderImage
🎯 What it does: A pluggable TryOn-Refiner module is proposed to reconstruct details of virtual try-on results, improving the accuracy of image details.
Tune-Your-Style: Intensity-tunable 3D Style Transfer with Gaussian Splatting
Yian Zhao (Peking University), Jie Chen (Peking University)
Image TranslationGenerationDiffusion modelGaussian SplattingPoint Cloud
🎯 What it does: A 3D style transfer framework called Tune-Your-Style is proposed, which allows adjustable style intensity. It utilizes Gaussian neurons and a learnable style regulator to achieve controllable style injection into 3D Gaussian projections, and enhances multi-view consistency through two-stage optimization and cross-view alignment.
Turbo2K: Towards Ultra-Efficient and High-Quality 2K Video Synthesis
Jingjing Ren (Hong Kong University of Science and Technology), Lei Zhu (South China University of Technology)
GenerationData SynthesisComputational EfficiencyKnowledge DistillationTransformerDiffusion modelAuto EncoderVideo
🎯 What it does: The Turbo2K framework is proposed to achieve high-quality 2K video synthesis and significantly improve inference speed.
TurboReg: TurboClique for Robust and Efficient Point Cloud Registration
Shaocheng Yan (Wuhan University), Jiayuan Li (Wuhan University)
RecognitionAutonomous DrivingOptimizationPoint Cloud
🎯 What it does: A TurboReg point cloud registration method is proposed, achieving fast and robust matching pair identification through TurboClique and Pivot-Guided Search.
TurboTrain: Towards Efficient and Balanced Multi-Task Learning for Multi-Agent Perception and Prediction
Zewei Zhou (University of California, Los Angeles), Jiaqi Ma (University of California, Los Angeles)
Autonomous DrivingComputational EfficiencyPoint CloudSequential
🎯 What it does: This paper presents TurboTrain, an efficient and balanced multi-task learning framework for end-to-end training of multi-agent perception and prediction.
TurboVSR: Fantastic Video Upscalers and Where to Find Them
Zhongdao Wang (Huawei Noah's Ark Lab), Wenbo Li (Huawei Noah's Ark Lab)
RestorationSuper ResolutionDiffusion modelAuto EncoderImageVideo
🎯 What it does: A highly efficient diffusion model video super-resolution framework TURBOVSR is proposed and implemented, which significantly improves inference speed while maintaining detail quality.
TWIST & SCOUT: Grounding Multimodal LLM-Experts by Forget-Free Tuning
Aritra Bhowmik (University of Amsterdam), Cees G. M. Snoek (University of Amsterdam)
RecognitionObject DetectionSegmentationTransformerLarge Language ModelMixture of ExpertsImageTextMultimodalityChain-of-Thought
🎯 What it does: Proposes the TWIST & SCOUT framework, which gradually fine-tunes a pre-trained multimodal large language model through dual experts to acquire visual localization capabilities while maintaining existing image-language understanding.
Two Losses, One Goal: Balancing Conflict Gradients for Semi-supervised Semantic Segmentation
Rui Sun (Shenzhen International Graduate School Tsinghua University), Yuan Wang (University of Science and Technology of China)
SegmentationOptimizationKnowledge DistillationImage
🎯 What it does: This paper studies a Pareto optimization strategy for semi-supervised semantic segmentation, dynamically balancing supervised and unsupervised gradients to address the gradient conflict issue.
U-ViLAR: Uncertainty-Aware Visual Localization for Autonomous Driving via Differentiable Association and Registration
Xiaofan Li (Baidu Inc), Jun Wang (Baidu Inc)
Pose EstimationAutonomous DrivingContrastive LearningSimultaneous Localization and MappingImage
🎯 What it does: Proposes U-ViLAR, a framework for uncertain visual localization that combines association and registration;
UAVScenes: A Multi-Modal Dataset for UAVs
Sijie Wang (Nanyang Technological University), Wee Peng Tay (Nanyang Technological University)
RecognitionSegmentationPose EstimationDepth EstimationSimultaneous Localization and MappingImageMultimodalityPoint CloudBenchmark
🎯 What it does: This paper presents UAVScenes—a large-scale, multi-modal drone dataset that provides frame-level image and LiDAR point cloud semantic annotations, 6 degrees of freedom (DoF) poses, and 3D maps, and conducts benchmark experiments for multiple tasks (image/point cloud semantic segmentation, depth estimation, localization, scene recognition, NVS, etc.).
UDC-VIT: A Real-World Video Dataset for Under-Display Cameras
Kyusu Ahn (Samsung Display Co., Ltd.), Jaejin Lee (Seoul National University)
RecognitionRestorationConvolutional Neural NetworkVideo
🎯 What it does: This paper designs a dual-camera system based on a beam splitter and aligns it with DFT to construct the first real-world UDC video dataset, UDC-VIT, and evaluates various deep learning models based on this dataset.
UINavBench: A Framework for Comprehensive Evaluation of Interactive Digital Agents
Harsh Agrawal (Apple), Alexander Toshev (Apple)
Large Language ModelAgentic AIVision Language ModelImageBenchmark
🎯 What it does: An online UI navigation benchmark for iOS, UINavBench, has been constructed, covering 36 applications and 116 tasks, and a task classification method along with a complete task validation process has been proposed.
UIP2P: Unsupervised Instruction-based Image Editing via Edit Reversibility Constraint
Enis Simsar (ETH Zurich), Federico Tombari (Technical University of Munich)
Image TranslationGenerationLarge Language ModelDiffusion modelImageText
🎯 What it does: This paper proposes an unsupervised instruction-driven image editing method called UIP2P, which utilizes the Edit Reversibility Constraint (ERC) to achieve consistency between forward and reverse editing in a single-step training process.
UIPro: Unleashing Superior Interaction Capability For GUI Agents
Hongxin Li (University of Chinese Academy of Sciences), Zhaoxiang Zhang (University of Chinese Academy of Sciences)
Robotic IntelligenceTransformerSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: A large-scale dataset for GUI understanding and interaction has been constructed, and a unified action space has been established, resulting in the training of a general GUI agent UIPro that can perform clicks, inputs, scrolling, and other operations across multiple platforms.
UKBOB: One Billion MRI Labeled Masks for Generalizable 3D Medical Image Segmentation
Emmanuelle Bourigault (University of Oxford), Abdullah Hamdi (University of Oxford)
SegmentationDomain AdaptationTransformerSupervised Fine-TuningImageBiomedical DataMagnetic Resonance ImagingBenchmark
🎯 What it does: The UKBOB dataset has been constructed, which is the largest voxel-level segmentation dataset of MRI images from the UK Biobank, and a foundational model called Swin-BOB based on Swin-UNetr has been trained.
ULTHO: Ultra-Lightweight yet Efficient Hyperparameter Optimization in Deep Reinforcement Learning
Mingqi Yuan (Hong Kong Polytechnic University), Wenjun Zeng (Ningbo Institute of Digital Twin)
OptimizationHyperparameter SearchReinforcement LearningTabular
🎯 What it does: ULTHO is proposed, a lightweight framework for hyperparameter optimization (HPO) conducted in a single training run.
Ultra High-Resolution Image Inpainting with Patch-Based Content Consistency Adapter
Jianhui Zhang (University of Electronic Science and Technology of China), Shuaicheng Liu (University of Electronic Science and Technology of China)
RestorationDiffusion modelImage
🎯 What it does: Proposes the Patch-Adapter two-stage lightweight adaptation framework, achieving 4K+ resolution text-guided image inpainting, addressing issues of content consistency and detail loss at high resolutions.
Ultra-Precision 6DoF Pose Estimation Using 2-D Interpolated Discrete Fourier Transform
Guowei Shi (Shanghai Jiao Tong University), Peisen Huang (Shanghai Jiao Tong University)
Pose EstimationImage
🎯 What it does: Utilizing 2D Interpolated Discrete Fourier Transform (2D-IpDFT) for precise estimation of frequency and phase in periodic pattern images, achieving ultra-high precision measurement of 6 degrees of freedom (6DoF) pose.
UMDATrack: Unified Multi-Domain Adaptive Tracking Under Adverse Weather Conditions
Siyuan Yao (Sun Yat-sen University), Xiaochun Cao (University of Chinese Academy of Sciences)
Object TrackingDomain AdaptationTransformerDiffusion modelVideo
🎯 What it does: A unified multi-domain adaptive tracking framework UMDATrack is proposed to maintain high-quality target state prediction under various adverse weather conditions.
Unbiased Missing-modality Multimodal Learning
Ruiting Dai (University of Electronic Science and Technology of China), Tao He (University of Electronic Science and Technology of China)
RestorationTransformerDiffusion modelScore-based ModelVideoTextMultimodalityStochastic Differential EquationAudio
🎯 What it does: This paper proposes a multi-stage bidirectional diffusion network (MD²) for unbiased recovery of missing modalities in multimodal learning.
Unbiased Region-Language Alignment for Open-Vocabulary Dense Prediction
Yunheng Li (Nankai University), Ming-Ming Cheng (Nankai University)
Object DetectionSegmentationVision Language ModelContrastive LearningImage
🎯 What it does: The DenseVLM framework is proposed to address the foreground bias problem in VLM for dense prediction through unsupervised region-text alignment, and it can directly replace the original VLM to enhance open vocabulary dense prediction performance.
Uncalibrated Structure from Motion on a Sphere
Jonathan Ventura (California Polytechnic State University), Fredrik Kahl (Chalmers University of Technology)
Pose EstimationOptimizationSimultaneous Localization and MappingOptical FlowVideo
🎯 What it does: A calibration-free structure from motion (SFM) method for spherical camera motion is proposed, utilizing a single unknown focal length for global reconstruction.
Uncertainty-Aware Diffusion-Guided Refinement of 3D Scenes
Sarosij Bose (University of California, Riverside), Amit K. Roy-Chowdhury (University of California, Riverside)
RestorationSegmentationGenerationDepth EstimationLarge Language ModelDiffusion modelVideoPoint Cloud
🎯 What it does: This paper presents UAR-Scenes, a diffusion-guided 3D scene refinement pipeline based on uncertainty perception, which can further enhance the rendering quality of new viewpoints on rough point clouds generated from single-view to 3D reconstruction models (such as Flash3D), especially for reasonable completion of unobserved areas.
Uncertainty-Aware Gradient Stabilization for Small Object Detection
Huixin Sun (Beihang University), Baochang Zhang (Beihang University)
Object DetectionImage
🎯 What it does: This paper proposes a gradient stabilization framework based on uncertainty perception, UGS, aimed at addressing the convergence difficulties caused by unstable localization gradients in small object detection.
Uncertainty-Driven Expert Control: Enhancing the Reliability of Medical Vision-Language Models
Xiao Liang (Xidian University), Tat-Seng Chua (National University of Singapore)
RecognitionRetrievalTransformerVision Language ModelMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: The Expert-CFG framework is proposed, which enhances the reliability of medical vision-language models without training by utilizing uncertainty estimation and expert annotations.
Uncover Treasures in DCT: Advancing JPEG Quality Enhancement by Exploiting Latent Correlations
Jing Yang (Beihang University), Minglang Qiao (Beihang University)
RestorationCompressionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an AJQE method for directly enhancing the quality of JPEG images in the DCT domain, avoiding the high computational cost associated with traditional pixel domain decoding.
Understanding Co-speech Gestures in-the-wild
Sindhu B Hegde (University of Oxford), Andrew Zisserman (University of Oxford)
RecognitionRetrievalConvolutional Neural NetworkTransformerContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: Proposes the JEGAL framework, which learns a joint embedding space for gestures, speech, and text modalities, and evaluates it on three new tasks (gesture retrieval, gesture-word localization, and gesture-based active speaker detection).
Understanding Flatness in Generative Models: Its Role and Benefits
Taehwan Lee (Ulsan National Institute of Science and Technology), Sung Whan Yoon (Ulsan National Institute of Science and Technology)
GenerationDiffusion modelImage
🎯 What it does: This study investigates the flat minima of diffusion models and explores their impact on generation quality, exposure bias, and quantization robustness.
Understanding Museum Exhibits using Vision-Language Reasoning
Ada-Astrid Balauca (INSAIT), Luc Van Gool (INSAIT)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: A large-scale multilingual museum exhibit visual question answering dataset, MUSEUM-65, was constructed, and the BLIP and LLaVA visual language models were fine-tuned on it to complete multi-task evaluations.
Understanding Personal Concept in Open-Vocabulary Semantic Segmentation
Sunghyun Park (Qualcomm AI Research), Fatih Porikli (Qualcomm AI Research)
SegmentationPrompt EngineeringVision Language ModelImage
🎯 What it does: A framework is proposed for personalized extension of the Open Vocabulary Semantic Segmentation (OVSS) model, which achieves 'my object' segmentation through a small number of images and masks;
Underwater Visual SLAM with Depth Uncertainty and Medium Modeling
Rui Liu (Zhejiang University), Yi Yang (Zhejiang University)
Object TrackingPose EstimationDepth EstimationSimultaneous Localization and MappingOptical FlowImageVideo
🎯 What it does: A framework for underwater visual SLAM called DUV-SLAM is proposed, which combines depth uncertainty quantification and light scattering medium modeling to enhance localization and reconstruction performance.
Unfolding-Associative Encoder-Decoder Network with Progressive Alignment for Pansharpening
Shijie Fang (Northwestern Polytechnical University), Hongping Gan (Northwestern Polytechnical University)
Image TranslationRestorationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a lightweight deep unfolded joint encoding-decoding network (UED-Net) that achieves high-quality fusion of remote sensing images by recursively encoding multi-level spatial-spectral degraded features and aggregating information at various stages.
UniCombine: Unified Multi-Conditional Combination with Diffusion Transformer
Haoxuan Wang (Fudan University), Yabiao Wang (Zhejiang University)
GenerationData SynthesisTransformerDiffusion modelRectified FlowImage
🎯 What it does: The UniCombine framework is proposed to achieve controllable image generation with multiple conditions (text, spatial graph, subject graph).
UniConvNet: Expanding Effective Receptive Field while Maintaining Asymptotically Gaussian Distribution for ConvNets of Any Scale
Yuhao Wang (Xi'an Jiaotong University), Wei Xi (Xi'an Jiaotong University)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a three-layer receptive field aggregation module (RFA), which achieves a large effective receptive field (ERF) within small convolutions through a hierarchical Amplifier-Discriminator combination while maintaining an Asymptotic Gaussian Distribution (AGD), thereby constructing a unified convolutional network, UniConvNet, that can be used across scales.
UniDxMD: Towards Unified Representation for Cross-Modal Unsupervised Domain Adaptation in 3D Semantic Segmentation
Zhengyin Liang (Beijing Jiaotong University), Hua Huang (Beijing Jiaotong University)
SegmentationDomain AdaptationAutonomous DrivingMultimodalityPoint Cloud
🎯 What it does: A unified discrete representation framework UniDxMD is proposed for cross-modal unsupervised domain adaptation in 3D semantic segmentation, addressing both modality bias and domain distribution differences.
UniEgoMotion: A Unified Model for Egocentric Motion Reconstruction, Forecasting, and Generation
Chaitanya Patel (Stanford University), Ehsan Adeli (Panasonic Holdings Corporation)
GenerationPose EstimationTransformerDiffusion modelVideo
🎯 What it does: Proposes UniEgoMotion, a unified conditional diffusion-based model capable of viewpoint motion reconstruction, prediction, and generation;
Unified Adversarial Augmentation for Improving Palmprint Recognition
Jianlong Jin (Hefei University of Technology), Yunsheng Wu (Hefei University of Technology)
RecognitionAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A unified adversarial augmentation framework (UAA) is proposed, which generates adversarial palm images with both geometric distortion and texture degradation through spatial transformation and identity-preserving generative networks, thereby enhancing the robustness of palm recognition models on complex samples.
Unified Category-Level Object Detection and Pose Estimation from RGB Images using 3D Prototypes
Tom Fischer (Saarland University), Eddy Ilg (University of Technology Nuremberg)
Object DetectionPose EstimationTransformerContrastive LearningImage
🎯 What it does: A unified RGB single-stage model is designed, using neural grid prototypes to simultaneously achieve category-level object detection and pose estimation.
Unified Multi-Agent Trajectory Modeling with Masked Trajectory Diffusion
Songru Yang (Beihang University), Zhengxia Zou (Beihang University)
Mixture of ExpertsDiffusion modelTime Series
🎯 What it does: This paper proposes UniMTD, a unified approach for multi-agent trajectory prediction, filling, and recovery, utilizing a decoupled masking diffusion framework.
Unified Multimodal Understanding via Byte-Pair Visual Encoding
Wanpeng Zhang (Peking University), Zongqing Lu (Peking University)
TransformerLarge Language ModelVision Language ModelGenerative Adversarial NetworkImageTextMultimodality
🎯 What it does: This work proposes a visual discretization method based on Byte Pair Encoding (BPE), which quantizes images into visual tokens and inputs these visual tokens along with text tokens into the same Transformer for cross-modal modeling.
Unified Open-World Segmentation with Multi-Modal Prompts
Yang Liu (Zhejiang University), Chunhua Shen (Zhejiang University)
SegmentationPrompt EngineeringImageVideoTextMultimodality
🎯 What it does: A unified open-world segmentation model COSINE is proposed, capable of performing various tasks such as semantic, instance, panoptic, referring, and video object segmentation through text and image multimodal prompts.
Unified Video Generation via Next-Set Prediction in Continuous Domain
Zhanzhou Feng (Peking University), Shiliang Zhang (Peking University)
GenerationData SynthesisTransformerAuto EncoderImageVideoTextMultimodality
🎯 What it does: This paper proposes a unified continuous domain autoregressive model that achieves high-quality generation of images and videos through a next-set prediction strategy.
UniFuse: A Unified All-in-One Framework for Multi-Modal Medical Image Fusion Under Diverse Degradations and Misalignments
Dayong Su (Kunming University of Science and Technology), Yu Liu (Hefei University of Technology)
RestorationSegmentationTransformerPrompt EngineeringMultimodalityBiomedical DataMagnetic Resonance ImagingComputed TomographyPositron Emission Tomography
🎯 What it does: The UniFuse framework is proposed to achieve integrated processing of multimodal medical images in three tasks: distortion, misalignment, and fusion.
UniGlyph: Unified Segmentation-Conditioned Diffusion for Precise Visual Text Synthesis
Yuanrui Wang (Tsinghua University), Liu Lin (Baidu Inc.)
SegmentationGenerationData SynthesisSupervised Fine-TuningDiffusion modelImageText
🎯 What it does: Proposes the UniGlyph framework, which uses pixel-level text segmentation masks instead of rendered glyph images to construct a unified conditional input, achieving precise visual text generation.
UniGS: Modeling Unitary 3D Gaussians for Novel View Synthesis from Sparse-view Images
Jiamin Wu (International Digital Economy Academy), Lei Zhang (International Digital Economy Academy)
GenerationData SynthesisGaussian SplattingImage
🎯 What it does: The UniGS model is proposed, which uses a unified world space 3D Gaussian set for sparse view 3D reconstruction and novel view synthesis.
UniMLVG: Unified Framework for Multi-view Long Video Generation with Comprehensive Control Capabilities for Autonomous Driving
Rui Chen (Southeast University), Siyu Xia
GenerationAutonomous DrivingDiffusion modelVideo
🎯 What it does: A unified multi-view long video generation framework, UniMLVG, has been designed to generate high-quality multi-view driving videos based on reference frames, text, and 3D conditions.
UniOcc: A Unified Benchmark for Occupancy Forecasting and Prediction in Autonomous Driving
Yuping Wang (University of Michigan), Jiachen Li (University of California)
Autonomous DrivingTransformerOptical FlowPoint CloudBenchmark
🎯 What it does: UniOcc proposes a unified benchmark and toolkit for occupancy prediction and forecasting, integrating real data from nuScenes and Waymo, as well as simulated data from CARLA and OpenCOOD, generating 2D/3D occupancy grids, motion flows for each voxel, and multi-agent collaborative scenarios, while providing a complete training/testing pipeline, visualization, and evaluation tools.