ICCV 2025 Papers — Page 27
IEEE/CVF International Conference on Computer Vision · 2701 papers
ViT-Split: Unleashing the Power of Vision Foundation Models via Efficient Splitting Heads
Yifan Li (Michigan State University), Liu Ren (Bosch Research North America & Bosch Center for Artificial Intelligence)
Object DetectionSegmentationTransformerImage
🎯 What it does: This paper proposes a visual foundation model adapter named ViT-Split, aimed at enhancing multi-task downstream performance solely through two lightweight heads (Prior head and Task head) while keeping the original VFM parameters frozen.
VITAL: More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow
Ada Görgün (Max Planck Institute for Informatics), Jonas Fischer (Max Planck Institute for Informatics)
GenerationExplainability and InterpretabilityConvolutional Neural NetworkTransformerImage
🎯 What it does: A new feature visualization method called VITAL is proposed, which aligns the feature distribution of generated images with that of real images and combines the correlation scores of neural networks to achieve more understandable visualizations.
Vivid4D: Improving 4D Reconstruction from Monocular Video by Video Inpainting
Jiaxin Huang (Zhejiang University), Yiyi Liao (ByteDance)
RestorationGenerationDepth EstimationDiffusion modelVideo
🎯 What it does: This paper proposes the Vivid4D method, which utilizes video augmentation and video inpainting techniques to enhance the quality of 4D dynamic scene reconstruction from monocular videos.
VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks
Shiduo Zhang (Fudan University), Xipeng Qiu (Fudan University)
Robotic IntelligenceLarge Language ModelVision Language ModelVision-Language-Action ModelPoint CloudBenchmark
🎯 What it does: This paper proposes and implements VLABench, a large-scale language-conditioned robotic manipulation benchmark, which includes 100 task categories, over 2000 3D objects, and provides an automated data collection and evaluation framework.
VLDrive: Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving
Ruifei Zhang (Chinese University of Hong Kong), Guanbin Li (Sun Yat-sen University)
Autonomous DrivingTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: A lightweight multimodal LLM framework called VLDrive is proposed for natural language instruction-driven autonomous driving.
VLIPP: Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior
Xindi Yang (Monash University), Xu Jia (Dalian University of Technology)
Object DetectionGenerationData SynthesisTransformerLarge Language ModelVision Language ModelDiffusion modelOptical FlowVideoTextBenchmarkPhysics RelatedChain-of-Thought
🎯 What it does: This paper proposes a two-stage visual language physical prior-driven video generation framework (VLIPP), which first uses a Vision-Language Model (VLM) to plan coarse-grained physical motion trajectories, and then refines and generates physically feasible videos using a video diffusion model (VDM).
VLM4D: Towards Spatiotemporal Awareness in Vision Language Models
Shijie Zhou (University of California, Los Angeles), Achuta Kadambi (University of California, Los Angeles)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextBenchmark
🎯 What it does: This paper proposes the VLM4D benchmark, aimed at systematically evaluating the spatiotemporal reasoning capabilities of visual language models (VLM) in four dimensions (3D space + time) regarding motion, rotation, counting, and misjudgment. It constructs a test set through question-and-answer pairs annotated by both humans and LLMs; it also explores enhancement methods such as targeted SFT and 4D feature field reconstruction.
VLR-Driver: Large Vision-Language-Reasoning Models for Embodied Autonomous Driving
Fanjie Kong, Hongbin Sun
Autonomous DrivingTransformerLarge Language ModelReinforcement LearningVision Language ModelImageVideoMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes VLR-Driver, a hierarchical Spatiotemporal Chain-of-Thought visual-language reasoning model for closed-loop embodied autonomous driving, and constructs the corresponding VLRDriver dataset.
VLRMBench: A Comprehensive and Challenging Benchmark for Vision-Language Reward Models
Jiacheng Ruan (Shanghai Jiao Tong University), Yuzhuo Fu
TransformerLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper presents VLRMBench, a visual-language reward model evaluation benchmark that includes 12 tasks and 12,634 questions, aimed at comprehensively examining the model's process understanding, result judgment, and critical generation capabilities.
VMBench: A Benchmark for Perception-Aligned Video Motion Generation
Xinran Ling (AMAP, Alibaba Group), Xiangxiang Chu (AMAP, Alibaba Group)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVideoBenchmark
🎯 What it does: Developed the VMBench video motion assessment benchmark, which includes five-dimensional human perception-aligned motion quality metrics and a large-scale diverse motion prompt library.
VMem: Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory
Runjia Li (University of Oxford), Tomas Jakab (University of Oxford)
GenerationData SynthesisComputational EfficiencyDiffusion modelImageVideo
🎯 What it does: This paper proposes a memory module based on surfels, called VMem, to efficiently retrieve and utilize past views in interactive video generation, starting from a single image to generate long-term, interactive 3D scenes.
VOccl3D: A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions
Yash Garg (University of California), Amit Roy-Chowdhury (University of California)
Object DetectionPose EstimationSupervised Fine-TuningGaussian SplattingVideoBenchmark
🎯 What it does: This paper presents VOccl3D, a large-scale, highly realistic synthetic video dataset designed for training and evaluating 3D human pose and shape estimation algorithms in significantly occluded scenes. It also fine-tunes existing models such as CLIFF, BEDLAM-CLIFF, and YOLO11 to validate their robustness under occlusion conditions.
VoiceCraft-Dub: Automated Video Dubbing with Neural Codec Language Models
Kim Sung-Bin (POSTECH), David Harwath (University of Texas at Austin)
GenerationData SynthesisTransformerLarge Language ModelVideoTextMultimodalityAudio
🎯 What it does: Developed VoiceCraft-Dub, an automatic video dubbing system that combines text, source voice, and target video facial features.
VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction
Martin de La Gorce (Microsoft), Antonio Criminisi (Microsoft)
GenerationData SynthesisConvolutional Neural NetworkGaussian SplattingVideoPoint Cloud
🎯 What it does: A method is proposed that can predict 3D Gaussian sparse point clouds in real-time from a single camera video stream, capable of generating realistic and stable 3D portraits from new perspectives while maintaining the authenticity of the input view (consistency between the input image and the reconstructed image).
VolumetricSMPL: A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions
Marko Mihajlovic (ETH Zurich), Siyu Tang (UC Berkeley)
Pose EstimationComputational EfficiencyPoint CloudMesh
🎯 What it does: This paper presents VolumetricSMPL, a volumetric model that represents human body shapes as Signed Distance Fields (SDF), supporting efficient and differentiable interactions between the human body and the environment.
VoteSplat: Hough Voting Gaussian Splatting for 3D Scene Understanding
Minchao Jiang (Xidian University), Liang Zhang
Object DetectionSegmentationRetrievalGaussian SplattingPoint Cloud
🎯 What it does: VoteSplat is proposed, a 3D scene understanding framework that combines Hough voting with 3D Gaussian Splatting. It utilizes 2D votes generated by SAM to guide the offset vectors in Gaussian primitives, achieving instance segmentation at the point cloud level, open vocabulary localization, and click-based 3D object editing.
VOVTrack: Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking
Zekun Qian (Tianjin University), Wei Feng (Tianjin University)
Object DetectionObject TrackingKnowledge DistillationPrompt EngineeringVision Language ModelContrastive LearningVideo
🎯 What it does: This paper presents VOVTrack, an end-to-end framework for open vocabulary multi-object tracking (OVMOT) that integrates three sub-tasks: object detection, classification, and association.
VoxelKP: A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data
Jian Shi (King Abdullah University of Science and Technology), Peter Wonka (King Abdullah University of Science and Technology)
Pose EstimationAutonomous DrivingConvolutional Neural NetworkPoint Cloud
🎯 What it does: A fully sparse network VoxelKP is proposed for 3D human keypoint estimation in single-frame LiDAR point clouds.
Voyaging into Perpetual Dynamic Scenes from a Single View
Fengrui Tian (University of Pennsylvania), Rene Vidal (University of Pennsylvania)
GenerationData SynthesisDiffusion modelVideoPoint Cloud
🎯 What it does: A system called DynamicVoyager is proposed, which generates 3D consistent, infinitely dynamic scenes from a single view (image or fixed-angle video) and can produce continuous dynamic content under any flying camera trajectory.
VPO: Aligning Text-to-Video Generation Models with Prompt Optimization
Jiale Cheng (Tsinghua University), Minlie Huang (Tsinghua University)
GenerationOptimizationSafty and PrivacyLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVideoText
🎯 What it does: This paper proposes the VPO framework, addressing the prompt optimization issue for text-to-video generation models. It constructs an alignment mechanism based on the three principles of 'harmless, accurate, and useful' and achieves safe, accurate, and helpful prompt generation through a two-stage process (principle-driven SFT and multi-feedback preference optimization).
VPR-Cloak: A First Look at Privacy Cloak Against Visual Place Recognition
Shuting Dong (Tsinghua University), Chun Yuan (Tsinghua University)
RecognitionOptimizationSafty and PrivacyTransformerContrastive LearningImage
🎯 What it does: In the visual place recognition (VPR) system, a privacy protection mechanism called VPR-Cloak is proposed, which generates imperceptible perturbations to suppress the identification of the opponent's model location information.
VQ-SGen: A Vector Quantized Stroke Representation for Creative Sketch Generation
Jiawei Wang (University of Edinburgh), Changjian Li (ShanghaiTech University)
GenerationData SynthesisTransformerAuto EncoderImage
🎯 What it does: This paper proposes VQ-SGen, a creative sketch generation framework that utilizes vector quantization stroke representation and autoregressive Transformers.
VQ-VLA: Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers
Yating Wang (Shanghai Research Institute for Intelligent Autonomous Systems), Tong He (Shanghai Research Institute for Intelligent Autonomous Systems)
GenerationOptimizationRobotic IntelligenceConvolutional Neural NetworkVision-Language-Action ModelMultimodality
🎯 What it does: Trained and deployed a vector quantization-based action tokenizer, embedding it into the Vision-Language-Action model, significantly improving the generation speed and continuity of action sequences.
VRBench: A Benchmark for Multi-Step Reasoning in Long Narrative Videos
Jiashuo Yu (Shanghai Artificial Intelligence Laboratory), Limin Wang (Nanjing University)
Large Language ModelVision Language ModelVideoTextBenchmark
🎯 What it does: This paper presents VRBench, a multi-step reasoning benchmark for long narrative videos, which collects 960 long videos and collaboratively annotates each video with 8-10 multi-step question-answer pairs and 25,106 timestamped reasoning steps, designing a multi-stage evaluation process from the result layer to the process layer.
VRM: Knowledge Distillation via Virtual Relation Matching
Weijia Zhang (Shanghai Jiao Tong University), Chao Ma (The University of Sydney)
Object DetectionKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: A knowledge distillation framework through Virtual Relation Matching (VRM) is proposed, which constructs a dense relation graph using generated virtual views, and subsequently transfers the student model through a pruning method.
VSC: Visual Search Compositional Text-to-Image Diffusion Model
Do Huu Dat (VinUniversity), Tae-Hyun Oh (KAIST)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: The VSC method is proposed, which splits composite text prompts into individual attribute-object pairs, generates visual prototype images separately, and then fuses visual features with text embeddings to enhance the attribute binding effect of the diffusion model.
VSP: Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs
Qiucheng Wu (University of California Santa Barbara), Shiyu Chang (University of California Santa Barbara)
Large Language ModelPrompt EngineeringImageTextMultimodalityBenchmark
🎯 What it does: This work proposes the VSP (Visual Spatial Planning) benchmark for evaluating and diagnosing the capabilities of multimodal large language models (MLLM) in visual spatial planning tasks, which includes a main planning task and four fine-grained sub-tasks.
VSRM: A Robust Mamba-Based Framework for Video Super-Resolution
Dinh Phu Tran (Korea Advanced Institute of Science and Technology), Daeyoung Kim (Korea Advanced Institute of Science and Technology)
RestorationSuper ResolutionOptical FlowVideo
🎯 What it does: A video super-resolution framework VSRM based on Mamba is proposed, which enhances image quality through multi-scale spatiotemporal modeling and specialized loss in the frequency domain.
VSSD: Vision Mamba with Non-Causal State Space Duality
Yuheng Shi (University of Sydney), Chang Xu (University of Sydney)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper researches and implements a non-causal state space model for visual tasks, VSSD, improving the original SSD to NC-SSD and constructing an efficient visual network.
VTimeCoT: Thinking by Drawing for Video Temporal Grounding and Reasoning
Jinglei Zhang (Shanghai Jiao Tong University), Chao Ma (Shanghai Jiao Tong University)
SegmentationRetrievalTransformerLarge Language ModelVision Language ModelVideoMultimodalityChain-of-Thought
🎯 What it does: Proposes the VTimeCoT framework, which utilizes a visual progress bar and efficient highlighting tools, enabling multimodal large language models (MLLM) to perform visual temporal chain reasoning in videos, thereby achieving video temporal localization and reasoning tasks.
Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection
Dat Nguyen (University of Luxembourg), Djamila Aouada (University of Luxembourg)
ClassificationData SynthesisTransformerVideo
🎯 What it does: The FakeSTormer framework is proposed, which captures spatial and temporal vulnerabilities in a fine-grained manner through multi-task learning, enhancing the generalization ability of deepfake video detection.
WalkVLM: Aid Visually Impaired People Walking by Vision Language Model
Zhiqiang Yuan (Tencent), Jinchao Zhang (Tencent)
Robotic IntelligenceTransformerLarge Language ModelVision Language ModelVideoTextChain-of-Thought
🎯 What it does: This paper constructs a blind walking assistance dataset (WAD) consisting of 12k video-annotation pairs and proposes the WalkVLM model, which provides timely, concise, and informative walking reminders and Q&A for visually impaired individuals using hierarchical planning and temporal adaptive prediction based on chain reasoning.
WarpHE4D: Dense 4D Head Map toward Full Head Reconstruction
Jongseob Yun (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
SegmentationGenerationDepth EstimationContrastive LearningImageMesh
🎯 What it does: Proposes the WarpHE4D framework to achieve full head 3D reconstruction and pixel-level UV coordinate prediction from a single image.
Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation
Xi Yu (Xi'an Jiaotong University), Jian Sun (Xi'an Jiaotong University)
Image TranslationGenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes a distribution transformation method based on Wasserstein distance, WSDT, for achieving image style transfer in diffusion models without training.
Wave-MambaAD: Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection
Qiao Zhang (China University of Petroleum), Kai Xu (China University of Petroleum)
Anomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: A multi-class unsupervised anomaly detection framework called Wave-MambaAD based on wavelet transform and state space model is proposed, which can detect both subtle anomalies and large-scale anomalies simultaneously.
WAVE: Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image
Jiwoo Park (Yonsei University), Seong Jae Hwang (Yonsei University)
GenerationData SynthesisDiffusion modelAuto EncoderGaussian SplattingImage
🎯 What it does: A training-free WAVE method is proposed, which achieves scene-level novel view synthesis through perspective-guided 3D warping using a single image, significantly enhancing view consistency.
Wavelet Policy: Lifting Scheme for Policy Learning in Long-Horizon Tasks
Hao Huang (New York University Abu Dhabi), Yi Fang (New York University Abu Dhabi)
Autonomous DrivingRobotic IntelligenceTransformerReinforcement LearningMultimodalityTime SeriesSequential
🎯 What it does: A learnable wavelet strategy framework based on a wavelet lifting scheme is proposed for policy learning in long-term tasks.
WaveMamba: Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection
Haodong Zhu (Beihang University), Baochang Zhang (Beihang University)
Object DetectionSupervised Fine-TuningImageMultimodality
🎯 What it does: The WaveMamba scheme is proposed, which maps RGB and IR features to the frequency domain through discrete wavelet transform, utilizing low-frequency information fusion and high-frequency enhancement, and combines it with an improved YOLOv8 detection head to achieve RGB-IR fusion object detection.
Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning
Yafei Zhang (Kunming University of Science and Technology), Jie Wen (Harbin Institute of Technology)
RecognitionRetrievalConvolutional Neural NetworkContrastive LearningImageMultimodality
🎯 What it does: This paper proposes a weakly supervised visible-infrared person re-identification method that can establish cross-modal identity correspondence using only single-modal identity labels.
Weakly-Supervised Learning of Dense Functional Correspondences
Stefan Stojanov (Stanford University), Jiajun Wu (Stanford University)
Object DetectionSegmentationRepresentation LearningTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: A weakly supervised method is proposed to learn dense functional correspondences between objects of different categories, capable of capturing pixel-level correspondences of object parts that perform the same function.
WeaveSeg: Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation
Jiajia Li (Shenzhen University), Jing Qin (Hong Kong Polytechnic University)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: A deep learning framework named WeaveSeg is proposed for nuclear instance segmentation in microscopic images.
Web Artifact Attacks Disrupt Vision Language Models
Maan Qraitem (Boston University), Bryan A. Plummer (Boston University)
OptimizationAdversarial AttackPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes the Web Artifact attack, which misleads visual-language models (VLMs) using various visual decoys such as non-matching text, non-text graphics, and text-embedded graphics, and designs a complete process for retrieval, evaluation, and location optimization.
What Changed and What Could Have Changed? State-Change Counterfactuals for Procedure-Aware Video Representation Learning
Chi-Hsi Kung (Indiana University), Yi-Hsuan Tsai (Atmanity Inc.)
RetrievalRepresentation LearningLarge Language ModelContrastive LearningVideoText
🎯 What it does: By using the action state descriptions generated by large language models and corresponding counterfactuals (hypothetical incorrect or missing operation results) as supervision signals, a hierarchical contrastive learning framework is constructed, enabling the video encoder to capture the real changes and potential variations of actions in the scene, thus achieving a deeper understanding of programmatic videos.
What Changed? Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models
Lorenzo Baraldi (University of Modena and Reggio Emilia), Rita Cucchiara (University of Modena and Reggio Emilia)
Image TranslationObject DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningImageMultimodality
🎯 What it does: The DICE (DIfference Coherence Estimator) model is proposed, which first detects object-level differences between the original image and the edited image, and then evaluates the consistency of each difference with user instructions to assess the performance of instruction-driven image editing models.
What If: Understanding Motion Through Sparse Interactions
Stefan Andreas Baumann (Computer Vision Foundation), Björn Ommer (Computer Vision Foundation)
SegmentationGenerationPose EstimationTransformerMixture of ExpertsVideo
🎯 What it does: This paper proposes the Flow Poke Transformer (FPT), which directly predicts the multimodal probability distribution of local motion through sparse 'poke' interactions, applicable for motion generation, motion segmentation, and interactive animation.
What Makes for Text to 360-degree Panorama Generation with Stable Diffusion?
Jinhong Ni (Australian National University), Jing Zhang (Australian National University)
GenerationData SynthesisTransformerMixture of ExpertsDiffusion modelImage
🎯 What it does: This paper focuses on the generation of 360° panoramic images from text, fine-tuning the Stable Diffusion model with LoRA, and proposes a single-branch framework called UniPano that only adjusts the value and output matrices in the attention module, providing a mechanism explanation for panoramic structure learning.
What to Distill? Fast Knowledge Distillation with Adaptive Sampling
Byungchul Chae (Kyung Hee University), Seonyeong Heo (Kyung Hee University)
Knowledge DistillationConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: A fast knowledge distillation method based on adaptive sampling (KDAS) is proposed, which accelerates training by dynamically selecting samples with high knowledge quantity and high quality.
What we need is explicit controllability: Training 3D gaze estimator using only facial images
Tingwei Li (Hangzhou Dianzi University), Buyu Liu (Harbin Institute of Technology)
GenerationPose EstimationGaussian SplattingImage
🎯 What it does: This study proposes an unsupervised 3D gaze estimation method that generates realistic facial images and accurate pseudo-gaze labels from facial images using controllable 3D Gaussian rendering, thereby training a gaze estimator.
What You Have is What You Track: Adaptive and Robust Multimodal Tracking
Yuedong Tan (China Telecom), Zongwei Wu
Object TrackingTransformerMixture of ExpertsVideoMultimodality
🎯 What it does: A unified framework called FlexTrack is designed to handle both complete and missing modality scenarios in multi-modal tracking.
What's in a Latent? Leveraging Diffusion Latent Space for Domain Generalization
Xavier Thomas (Boston University), Deepti Ghadiyaram (Runway)
Domain AdaptationDiffusion modelImage
🎯 What it does: The paper proposes a method for pseudo-domain discovery using the latent space of pre-trained models (especially diffusion models), and concatenates these pseudo-domain features with standard classifier features to enhance generalization to unseen domains.
What's Making That Sound Right Now? Video-centric Audio-Visual Localization
Hahyeon Choi (Seoul National University), Nojun Kwak (Seoul National University)
Convolutional Neural NetworkContrastive LearningVideoMultimodalityBenchmarkAudio
🎯 What it does: A video-oriented audio-visual source localization benchmark AVATAR and a temporal-aware model TAVLO have been designed, supporting various real-world scenarios such as single sound source, mixed sound sources, multiple entities, and off-screen, achieving precise spatiotemporal audio source localization.
When Anchors Meet Cold Diffusion: A Multi-Stage Approach to Lane Detection
Bo-Lun Huang (National Yang Ming Chiao Tung University), Wen-Huang Cheng (National Taiwan University)
Autonomous DrivingConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: This paper proposes CDiffLane, a lane detection framework based on cold diffusion, which gradually refines lane predictions through a multi-step iterative approach starting from preset anchor points, and introduces resolution scheduling and a learnable α parameter to achieve fine-tuning from coarse to fine.
When and Where do Data Poisons Attack Textual Inversion?
Jeremy Styborski (Nanyang Technological University), Adams Wai-Kin Kong (Nanyang Technological University)
GenerationData SynthesisAdversarial AttackDiffusion modelImage
🎯 What it does: Analyzes the spatiotemporal learning bias of Textual Inversion (TI) in diffusion models under adversarial data poisoning attacks, and proposes a secure training scheme called Safe-Zone Training (SZT) based on JPEG compression, timestep limits, and loss masking to defend against various poisoning methods.
When Confidence Fails: Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation
Pan Liu (Central South University), Jinshi Liu (Central South University)
SegmentationOptimizationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A new pseudo-label selection method (Confidence Separable Learning, CSL) is proposed, which dynamically separates reliable and unreliable pixels by constructing a confidence distribution feature space and performing convex optimization; it also introduces Trusted Mask Perturbation to enhance contextual learning in low-confidence areas.
When Large Vision-Language Model Meets Large Remote Sensing Imagery: Coarse-to-Fine Text-Guided Token Pruning
Junwei Luo (Wuhan University), Yansheng Li (Wuhan University)
RecognitionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This paper proposes a text-guided token pruning method that combines dynamic image pyramids for efficient understanding of large-scale remote sensing images.
When Lighting Deceives: Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack
Hanqing Liu (Beihang University), Xingxing Wei (Beihang University)
OptimizationAdversarial AttackVision Language ModelImage
🎯 What it does: Generate illumination-aware adversarial examples to assess the robustness of Vision-Language models against lighting variations.
When Pixel Difference Patterns Meet ViT: PiDiViT for Few-Shot Object Detection
Hongliang Zhou (National University of Defense Technology), Li Liu (National University of Defense Technology)
Object DetectionTransformerImage
🎯 What it does: Proposes the PiDiViT model, which incorporates a pixel difference convolution fusion module (DCFM) and a multi-scale feature fusion module (MFFM) for few-shot object detection, improving the pre-trained features of ViT.
When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training
Yunwei Lan (University of Science and Technology of China), Dong Liu (University of Science and Technology of China)
Image TranslationRestorationContrastive LearningImage
🎯 What it does: This paper proposes a dehazing framework called DehazeSB based on the Schrödinger Bridge, which combines prompt learning and detail-preserving regularization to achieve optimal transport mapping from hazy images to clear images.
Where am I? Cross-View Geo-localization with Natural Language Descriptions
Junyan Ye (Sun Yat-Sen University), Weijia Li (Sun Yat-Sen University)
RetrievalExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A cross-perspective geographic localization task based on natural language descriptions is proposed, and a method for locating text-retrieved satellite images/OSM images is implemented.
Where, What, Why: Towards Explainable Driver Attention Prediction
Yuchen Zhou (Sun Yat-sen University), Chao Gou (Sun Yat-sen University)
Autonomous DrivingExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningImageVideoTextMultimodalityChain-of-Thought
🎯 What it does: An explainable driving attention prediction framework called LLada is proposed, which jointly predicts the focus area (Where), focus semantics (What), and attention reasons (Why).
Who Controls the Authorization? Invertible Networks for Copyright Protection in Text-to-Image Synthesis
Baoyue Hu (Chongqing University of Posts and Telecommunications), Bin Xiao (Inspur Data Technology Co., Ltd.)
GenerationData SynthesisSafty and PrivacyAuto EncoderContrastive LearningImage
🎯 What it does: A copyright protection framework based on reversible networks is proposed, which can prevent unauthorized personalized generation by embedding watermarks in the low-frequency domain while supporting authorized personalization and achieving copyright traceability.
Who is a Better Talker: Subjective and Objective Quality Assessment for AI-Generated Talking Heads
Yingjie Zhou (Shanghai Jiao Tong University), Guangtao Zhai (Shanghai Jiao Tong University)
GenerationData SynthesisTransformerVideoAudio
🎯 What it does: This paper constructs the largest AI-generated talking head (AGTH) quality assessment dataset THQA‑10K and proposes a novel objective evaluation method FSCD based on the first frame, Y-T slices, and audio-visual consistency.
Why LVLMs Are More Prone to Hallucinations in Longer Responses: The Role of Context
Ge Zheng (Sun Yat-sen University), Sibei Yang (ShanghaiTech University)
GenerationTransformerVision Language ModelContrastive LearningTextMultimodality
🎯 What it does: This study investigates the hallucination problem that arises in large visual language models during long text generation. It proposes an active hallucination instance induction-detection-suppression framework (HalTrapper) that induces hallucination instances, detects them based on two main hypotheses: attention similarity and image context integrity, and suppresses hallucinations during the decoding phase using contrastive context decoding.
Wide2Long: Learning Lens Compression and Perspective Adjustment for Wide-Angle to Telephoto Translation
Soumyadipta Banerjee (Indian Institute of Technology Kharagpur), Debashis Sen (Indian Institute of Technology Kharagpur)
Image TranslationDepth EstimationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A pipeline called Wide2Long is proposed to convert wide-angle images into telephoto images, simulating lens compression and perspective adjustment.
WikiAutoGen: Towards Multi-Modal Wikipedia-Style Article Generation
Zhongyu Yang (King Abdullah University of Science and Technology), Mohamed Elhoseiny (King Abdullah University of Science and Technology)
GenerationRetrievalTransformerLarge Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Developed the WikiAutoGen framework to achieve the automatic generation of multimodal (text + image) Wikipedia-style articles, and introduced a multi-perspective self-reflection mechanism.
WildSAT: Learning Satellite Image Representations from Wildlife Observations
Rangel Daroya (University of Massachusetts), Subhransu Maji (University of Massachusetts)
ClassificationSegmentationRetrievalRepresentation LearningContrastive LearningImageTextMultimodality
🎯 What it does: Using wildlife observation data along with multimodal information from satellite imagery, text, and environmental variables, the WildSAT method is proposed to perform contrastive learning on satellite image encoders.
WildSeg3D: Segment Any 3D Objects in the Wild from 2D Images
Yansong Guo (Xiamen University), Liujuan Cao (National University of Singapore)
Object DetectionSegmentationImagePoint Cloud
🎯 What it does: We propose WildSeg3D, a feed-forward 3D object segmentation framework based on 2D images, capable of real-time segmentation of arbitrary 3D objects without scene-specific training.
WINS: Winograd Structured Pruning for Fast Winograd Convolution
Cheonjun Park (Hankuk University of Foreign Studies), Won Woo Ro (Yonsei University)
OptimizationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A structured pruning method (WINS) is proposed for Winograd convolution, and further balanced pruning (WINS-B) and adaptive balanced pruning (WINS-AB) are introduced to balance speed and accuracy.
WIPES: Wavelet-based Visual Primitives
Wenhao Zhang (Nanjing University), Zhan Ma (Nanjing University)
Data SynthesisNeural Radiance FieldImage
🎯 What it does: A continuous differentiable visual primitive WIPES based on Morlet wavelets is proposed for the unified representation of multidimensional visual signals, and an efficient differentiable rasterizer is implemented;
WIR3D: Visually-Informed and Geometry-Aware 3D Shape Abstraction
Richard Liu (University of Chicago), Rana Hanocka (University of Chicago)
GenerationOptimizationDiffusion modelMesh
🎯 What it does: This paper proposes a method called WIR3D for abstracting visual and geometric information of 3D shapes using sparse 3D Bézier curves.
WonderPlay: Dynamic 3D Scene Generation from a Single Image and Actions
Zizhang Li (Stanford University), Jiajun Wu (Stanford University)
GenerationData SynthesisDiffusion modelImageVideoStochastic Differential Equation
🎯 What it does: We propose WonderPlay, a framework that generates dynamic 3D scenes from a single image and actions (such as gravity, wind fields, and point forces), supporting various physical materials (rigid bodies, elastic materials, fabrics, liquids, gases, particles, etc.).
WonderTurbo: Generating Interactive 3D World in 0.72 Seconds
Chaojun Ni (Peking University), Wenjun Mei (Peking University)
GenerationData SynthesisDiffusion modelGaussian SplattingImagePoint CloudOrdinary Differential Equation
🎯 What it does: The WonderTurbo framework is proposed, achieving real-time interactive 3D scene generation from a single image.
World4Drive: End-to-End Autonomous Driving via Intention-aware Physical Latent World Model
Yupeng Zheng (Institute of Automation, Chinese Academy of Sciences), Dongbin Zhao (Tsinghua University)
Autonomous DrivingTransformerVision Language ModelWorld ModelMultimodality
🎯 What it does: This paper proposes World4Drive, an intent-aware physical latent world model that achieves perception annotation-free end-to-end autonomous driving planning.
WorldScore: A Unified Evaluation Benchmark for World Generation
Haoyi Duan (Stanford University), Jiajun Wu (Stanford University)
GenerationData SynthesisDiffusion modelSimultaneous Localization and MappingOptical FlowImageVideoTextMultimodalityBenchmark
🎯 What it does: A unified world generation evaluation benchmark called WorldScore is proposed, along with a test set of 3000 examples covering static and dynamic, indoor and outdoor, and stylized scenes, to assess the capabilities of 3D/4D and video generation models in controllability, quality, and dynamic performance.
WSI-LLaVA: A Multimodal Large Language Model for Whole Slide Image
Yuci Liang (Shenzhen University), Linlin Shen (Shenzhen University)
ClassificationRecognitionSegmentationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringImageTextMultimodalityBiomedical DataBenchmark
🎯 What it does: The paper proposes the WSI-LLaVA framework and the WSI-Bench benchmark for multimodal understanding and diagnosis of whole slide images.
X-Capture: An Open-Source Portable Device for Multi-Sensory Learning
Samuel Clarke (Stanford University), Jiajun Wu (Stanford University)
GenerationRetrievalContrastive LearningMultimodalityPoint CloudAudio
🎯 What it does: Designed and implemented a low-cost open-source device X-Capture for synchronously collecting multi-sensory data such as RGBD, haptic, audio, and point clouds in field environments, and constructed a multimodal dataset containing 600 everyday objects and 3600 sampling points.
X-Dancer: Expressive Music to Human Dance Video Generation
Zeyuan Chen (University of California San Diego), Linjie Luo (ByteDance)
GenerationData SynthesisPose EstimationTransformerDiffusion modelVideoAudio
🎯 What it does: This paper proposes a zero-shot music-driven dance video generation framework called X-Dancer, which starts from a single static portrait. It utilizes a Transformer-diffusion joint model to automatically generate long-duration, music-synchronized, and diverse 2D pose sequences, and then achieves high-quality video synthesis through a diffusion model.
X-Fusion: Introducing New Modality to Frozen Large Language Models
Sicheng Mo (University of California), Yuheng Li (Adobe Research)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Proposes the X-Fusion framework, which utilizes a frozen large language model (LLM) and a specialized vision tower to achieve multimodal understanding and generation while maintaining the original language capabilities; based on this, a data-driven training strategy and an optional X-Fuse module are provided.
X-Prompt: Generalizable Auto-Regressive Visual Learning with In-Context Prompting
Zeyi Sun (Shanghai Jiao Tong University), Jiaqi Wang (Shanghai Jiao Tong University)
SegmentationGenerationDepth EstimationTransformerPrompt EngineeringVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: A pure autoregressive vision-language model, X-Prompt, has been designed and implemented, capable of performing various visual tasks such as image generation, editing, low-level processing, and dense perception by incorporating visual and textual prompts (in-context prompting) into the input.
X2-Gaussian: 4D Radiative Gaussian Splatting for Continuous-time Tomographic Reconstruction
Weihao Yu (Chinese University of Hong Kong), Yixuan Yuan (Chinese University of Hong Kong)
RestorationSegmentationGaussian SplattingImageVideoComputed Tomography
🎯 What it does: A continuous time 4D CT reconstruction framework based on dynamic radiative Gaussian scattering, called X-Gaussian, is proposed, which can directly recover a complete spatiotemporal continuous three-dimensional image from projection images.
X2I: Seamless Integration of Multimodal Understanding into Diffusion Transformer via Attention Distillation
Jian Ma (OPPO AI Center), Zhenyu Yang (OPPO AI Center)
GenerationData SynthesisKnowledge DistillationTransformerLarge Language ModelDiffusion modelImageVideoTextMultimodalityAudio
🎯 What it does: Proposes the X2I framework, which transfers the understanding capabilities of multimodal large language models to the Diffusion Transformer, enabling image generation and editing from multimodal inputs such as text, multiple languages, images, videos, and audio.
XTrack: Multimodal Training Boosts RGB-X Video Object Trackers
Yuedong Tan (University of Wurzburg), Radu Timofte (University of Wurzburg)
Object TrackingTransformerMixture of ExpertsVideoMultimodality
🎯 What it does: This paper proposes a cross-modal knowledge sharing framework based on Mixture of Modal Experts to address the issues of data sparsity and modality isolation in multi-modal visual tracking.
YOLO-Count: Differentiable Object Counting for Text-to-Image Generation
Guanning Zeng (Tsinghua University), Zhuowen Tu (UC San Diego)
Object DetectionGenerationConvolutional Neural NetworkVision Language ModelImage
🎯 What it does: We propose YOLO-Count, a differentiable open-vocabulary object counting model for precise control of object quantity in text-to-image generation.
YOLOE: Real-Time Seeing Anything
Ao Wang (Tsinghua University), Guiguang Ding (Tsinghua University)
Object DetectionSegmentationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A unified and efficient YOLOE model has been designed and implemented, capable of real-time detection and segmentation in open-source scenarios with text prompts, visual prompts, and no prompts.
You Are Your Own Best Teacher: Achieving Centralized-level Performance in Federated Learning under Heterogeneous and Long-tailed Data
Shanshan Yan (Xiamen University), Hanzi Wang (Xiamen University)
Federated LearningKnowledge DistillationContrastive LearningImage
🎯 What it does: This paper proposes FedYoYo, which addresses representation bias caused by heterogeneous and long-tailed data in federated learning through weak-strong enhanced self-teaching and distribution-aware Logit adjustment.
You Share Beliefs, I Adapt: Progressive Heterogeneous Collaborative Perception
Hao Si (University of Tokyo), Manabu Tsukada (University of Tokyo)
Object DetectionDomain AdaptationSupervised Fine-TuningMultimodality
🎯 What it does: A framework PHCP is designed and validated to adaptively tune the adapter during the inference phase using a small amount of unlabeled data, achieving heterogeneous collaborative perception.
You Think, You ACT: The New Task of Arbitrary Text to Motion Generation
Runqi Wang (Wuhan University), Zheng Wang (Wuhan University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningVideoText
🎯 What it does: A new task of Arbitrary Text to Motion is proposed, and a HUMANML3D++ dataset containing scene text and action labels is constructed.
Your Text Encoder Can Be An Object-Level Watermarking Controller
Naresh Kumar Devulapally (University at Buffalo), Vishnu Suresh Lokhande (University at Buffalo)
Object DetectionGenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelImageText
🎯 What it does: This paper proposes a pseudo-token-based implicit watermarking method that achieves watermark embedding for the entire image or selected objects during the generation process of the latent diffusion model (LDM) from text to image.
Zero-AVSR: Zero-Shot Audio-Visual Speech Recognition with LLMs by Learning Language-Agnostic Speech Representations
Jeong Hun Yeo (KAIST), Yong Man Ro (KAIST)
RecognitionTransformerLarge Language ModelVideoMultimodalityAudio
🎯 What it does: This paper proposes a zero-shot audio-video speech recognition framework called Zero-AVSR, which can recognize a language without any target language speech data.
Zero-Shot Composed Image Retrieval via Dual-Stream Instruction-Aware Distillation
Wenliang Zhong (University of Texas at Arlington), Junzhou Huang (Amazon)
RetrievalKnowledge DistillationLarge Language ModelContrastive LearningImageTextMultimodalityChain-of-Thought
🎯 What it does: This work proposes DistillCIR, which transfers the instruction-following capability of large language models to a lightweight projection-based retrieval model through dual-stream distillation, achieving unsupervised zero-shot composite image retrieval.
Zero-Shot Compositional Video Learning with Coding Rate Reduction
Heeseok Jung (Hanyang University), Eun-Sol Kim (Hanyang University)
RecognitionCompressionTransformerVision Language ModelVideo
🎯 What it does: A zero-shot compositional video understanding framework based on the decline of coding rate in information theory is proposed, utilizing a resampling transformer to decouple spatiotemporal features and achieve predictions of unknown compositional actions.
Zero-Shot Depth Aware Image Editing with Diffusion Models
Rishubh Parihar (Indian Institute of Science Bangalore), R. Venkatesh Babu (Indian Institute of Science Bangalore)
Image HarmonizationGenerationDepth EstimationDiffusion modelImageBenchmark
🎯 What it does: This paper proposes a zero-shot depth-aware image editing framework that can layer images and perform precise synthesis based on user-specified depth values.
Zero-shot Inexact CAD Model Alignment from a Single Image
Pattaramanee Arsomngern (VISTEC), Supasorn Suwajanakorn (VISTEC)
Object DetectionPose EstimationContrastive LearningImage
🎯 What it does: This paper proposes a zero-shot, scene-level pose annotation-free single-image CAD model alignment method that can solve the 9-DoF pose between retrieved approximate CAD models and target objects;
Zero-Shot Vision Encoder Grafting via LLM Surrogates
Kaiyu Yue (University of Maryland), Tom Goldstein (University of Maryland)
RecognitionCompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: This study proposes constructing a small language model surrogate and training a visual encoder on it, then directly integrating this encoder onto a large-scale LLM without additional training to achieve zero-shot visual language reasoning.
ZeroKey: Point-Level Reasoning and Zero-Shot 3D Keypoint Detection from Large Language Models
Bingchen Gong, Maks Ovsjanikov
RecognitionObject DetectionTransformerLarge Language ModelPrompt EngineeringImagePoint Cloud
🎯 What it does: A completely unsupervised, zero-shot 3D keypoint detection method called ZeroKey is proposed, which utilizes the pixel-level reasoning capability of a multimodal large language model (Molmo) to locate keypoints in multi-view images, and then maps and aggregates the 2D points into 3D space through back-projection and clustering.
ZeroStereo: Zero-shot Stereo Matching from Single Images
Xianqi Wang (Huazhong University of Science and Technology), Xin Yang (Huazhong University of Science and Technology)
GenerationData SynthesisDepth EstimationDiffusion modelImageBenchmark
🎯 What it does: A zero-shot matching pipeline called ZeroStereo is constructed to generate stereo images from a single image.
Zeroth-Order Fine-Tuning of LLMs in Random Subspaces
Ziming Yu (Beijing Normal University), Hua Huang (Beijing Normal University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: We propose SubZero, a zero-order optimization framework based on hierarchical low-rank random subspaces for fine-tuning large-scale LLMs, which avoids backpropagation and significantly reduces memory usage.
ZFusion: Efficient Deep Compositional Zero-shot Learning for Blind Image Super-Resolution with Generative Diffusion Prior
Alireza Esmaeilzehi (York University), Laleh Seyyed-Kalantari (York University)
RestorationSuper ResolutionDiffusion modelImage
🎯 What it does: A blind image super-resolution method based on diffusion models, ZFusion, is proposed, which utilizes synthetic zero-shot learning to estimate unknown degradations and embed them into the generation process.
ZIM: Zero-Shot Image Matting for Anything
Beomyoung Kim, Joonsang Yu
RestorationSegmentationTransformerImage
🎯 What it does: This paper proposes a zero-shot image matting model ZIM, which can generate high-precision fine-grained matting results without relying on manually annotated particle-level matting data.
ZipVL: Accelerating Vision-Language Models through Dynamic Token Sparsity
Yefei He (Zhejiang University), Bohan Zhuang (Zhejiang University)
Computational EfficiencyTransformerVision Language ModelImageVideo
🎯 What it does: The ZipVL framework is proposed, achieving efficient inference of large-scale vision-language models by dynamically allocating the proportion of important visual tokens.