ICCV 2025 Papers — Page 19
IEEE/CVF International Conference on Computer Vision · 2701 papers
PhysTwin: Physics-Informed Reconstruction and Simulation of Deformable Objects from Videos
Hanxiao Jiang (Columbia University), Yunzhu Li (Columbia University)
OptimizationRobotic IntelligenceVideoPhysics Related
🎯 What it does: A PhysTwin framework is constructed to utilize sparse video reconstruction for interactive physical digital twins, preserving geometric appearance while accurately simulating physical behavior.
Pi-GPS: Enhancing Geometry Problem Solving by Unleashing the Power of Diagrammatic Information
Junbo Zhao (Beijing Normal University), Hua Huang (Beijing Normal University)
Large Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Proposes the Pi-GPS framework, which utilizes image information to eliminate textual ambiguities in geometry problems, thereby improving the accuracy of geometric problem-solving.
Pinco: Position-induced Consistent Adapter for Diffusion Transformer in Foreground-conditioned Inpainting
Guangben Lu (Shanghai Jiao Tong University), Fangyuan Zou (Tencent)
RestorationGenerationTransformerDiffusion modelImage
🎯 What it does: A pluggable adapter called Pinco is proposed to achieve high-quality background filling in the diffusion transformer model under foreground conditions.
PINO: Person-Interaction Noise Optimization for Long-Duration and Customizable Motion Generation of Arbitrary-Sized Groups
Sakuya Ota (Institute of Science Tokyo), Ikuro Sato (Institute of Science Tokyo)
GenerationOptimizationDiffusion modelVideo
🎯 What it does: This paper proposes a training-agnostic framework called PINO for generating natural interactive actions of groups of characters of any size.
PixelStitch: Structure-Preserving Pixel-Wise Bidirectional Warps for Unsupervised Image Stitching
Hengzhe Jin (Beijing Jiaotong University), Yao Zhao (Beijing Jiaotong University)
Image TranslationOptimizationOptical FlowImage
🎯 What it does: We designed and implemented PixelStitch, an unsupervised image stitching framework based on bidirectional pixel-level distortion, capable of achieving high-quality stitching while maintaining structural integrity.
PixTalk: Controlling Photorealistic Image Processing and Editing with Language
Marcos V. Conde (University of Wurzburg), Radu Timofte (University of Wurzburg)
Image TranslationRestorationPrompt EngineeringMixture of ExpertsImageText
🎯 What it does: This paper presents PixTalk, a multi-task image processing model based on natural language prompts, capable of performing over 40 photography post-processing transformations at a speed of seconds on consumer GPUs for 12MP images.
PLA: Prompt Learning Attack against Text-to-Image Generative Models
Xinqi Lyu (Hong Kong Polytechnic University), Bin Xiao (Hong Kong Polytechnic University)
GenerationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes a gradient-driven prompt learning attack framework (PLA) for black-box text-to-image models, which generates NSFW generation prompts that can bypass prompt filters and post-hoc safety checkers by utilizing sensitive knowledge encoding and multimodal similarity loss.
PlaceIt3D: Language-Guided Object Placement in Real 3D Scenes
Ahmed Abdelreheem (King Abdullah University of Science and Technology), Guillermo Garcia-Hernando (Niantic Inc)
TransformerLarge Language ModelPoint CloudBenchmark
🎯 What it does: This paper proposes a language-driven 3D object placement task called PLACEIT3D, providing benchmarks and training datasets.
PLADIS: Pushing the Limits of Attention in Diffusion Models at Inference Time by Leveraging Sparsity
Kwanyoung Kim (Samsung Research), Byeongsu Sim (Samsung Research)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: During the inference phase of the diffusion model, the introduction of sparse attention (α‑Entmax) and a weighted compensation for dense attention enhance text alignment and image quality without the need for additional training or network evaluation.
PLAN: Proactive Low-Rank Allocation for Continual Learning
Xiequn Wang (Southern University of Science and Technology), Yu Zhang (Southern University of Science and Technology)
TransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes a continuous learning framework named PLAN, which utilizes low-rank adapters (LoRA) and actively allocates orthogonal subspaces to avoid interference between tasks.
Planar Affine Rectification from Local Change of Scale and Orientation
Yuval Nissan, Daniel Barath
OptimizationImage
🎯 What it does: A plane affine correction method based on local scale and direction changes is proposed, achieving efficient least squares solutions by combining linear constraints of scale and direction.
PlaneRAS: Learning Planar Primitives for 3D Plane Recovery
Fang Zhang (Beijing University of Posts and Telecommunications), Xiuzhuang Zhou (Beijing University of Posts and Telecommunications)
SegmentationDepth EstimationTransformerGaussian SplattingImage
🎯 What it does: This paper proposes the PlaneRAS framework, which utilizes a single image to explicitly reconstruct planar primitives in 3D space, aggregate them, and perform splatting to achieve 3D plane recovery and pixel-level plane mask prediction.
PlanGen: Towards Unified Layout Planning and Image Generation in Auto-Regressive Vision Language Models
Runze He (360 AI Research), Yuhui Yin (360 AI Research)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelImageText
🎯 What it does: PlanGen is a unified autoregressive visual language model capable of performing multiple tasks such as layout planning, layout-to-image generation, image layout understanding, and layout-based image editing within the same Transformer.
Player-Centric Multimodal Prompt Generation for Large Language Model Based Identity-Aware Basketball Video Captioning
Zeyu Xi (Beijing University of Technology), Changwen Chen (Hong Kong Polytechnic University)
GenerationTransformerLarge Language ModelPrompt EngineeringVideoMultimodality
🎯 What it does: A player-centered multimodal prompt generation network based on large language models is proposed to achieve identity-aware basketball video subtitle generation.
PLMP - Point-Line Minimal Problems for Projective SfM
Kim Kiehn (KTH Royal Institute of Technology), Kathlén Kohn (KTH Royal Institute of Technology)
Optimization
🎯 What it does: This paper systematically classifies and solves minimal problems for three-dimensional scenes composed of points and lines observed using uncalibrated pinhole cameras from complete multi-view perspectives, and provides the algebraic degree for each problem.
Plug-in Feedback Self-adaptive Attention in CLIP for Training-free Open-Vocabulary Segmentation
Zhixiang Chi (University of Toronto), Konstantinos Plataniotis
SegmentationTransformerContrastive LearningImage
🎯 What it does: The research objective is to enhance the spatial consistency and accuracy of CLIP in open vocabulary semantic segmentation through an untrained adaptive mechanism.
PlugMark: A Plug-in Zero-Watermarking Framework for Diffusion Models
Pengzhen Chen (Institute of Information Engineering, Chinese Academy of Sciences), Wu Liu (University of Science and Technology of China)
GenerationData SynthesisAdversarial AttackConvolutional Neural NetworkDiffusion modelImageMultimodality
🎯 What it does: This paper proposes PlugMark, a plug-and-play zero watermark framework for diffusion models, aimed at achieving unique identification and copyright tracking of model knowledge without compromising generation quality.
Point Cloud Self-supervised Learning via 3D to Multi-view Masked Learner
Zhimin Chen (Clemson University), Bing Li
ClassificationObject DetectionSegmentationRepresentation LearningTransformerAuto EncoderPoint Cloud
🎯 What it does: A self-supervised learning framework called Multiview-ML is proposed, which solely uses 3D point clouds. By projecting point clouds into multi-view 2D images at the 3D feature layer and utilizing an autoencoder for reconstruction, it achieves multi-modal feature learning.
PointGAC: Geometric-Aware Codebook for Masked Point Modeling
Abiao Li (Jiangxi University of Finance and Economics), Guofeng Mei (Fondazione Bruno Kessler)
SegmentationRepresentation LearningTransformerContrastive LearningPoint Cloud
🎯 What it does: A clustering-based self-supervised point cloud masking modeling method called PointGAC is proposed, which utilizes an online codebook-guided teacher-student framework to align masked region features through clustering, thereby learning more generalized representations.
PolarAnything: Diffusion-based Polarimetric Image Synthesis
Kailong Zhang (Beijing University of Posts and Telecommunications), Boxin Shi (Peking University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A PolarAnything based on diffusion models is proposed, which can generate high-quality, physically accurate polarized images using only a single RGB image.
Polarimetric Neural Field via Unified Complex-Valued Wave Representation
Chu Zhou (National Institute of Informatics), Imari Sato (National Institute of Informatics)
RestorationSuper ResolutionImage
🎯 What it does: A polarized neural field is proposed, which continuously reconstructs discrete polarization parameters (Total Intensity TI, Degree of Polarization DoP, Angle of Polarization AoP) through a unified complex wave representation method.
PolGS: Polarimetric Gaussian Splatting for Fast Reflective Surface Reconstruction
Yufei Han (Beijing University of Posts and Telecommunications), Zhanyu Ma (Peking University)
RestorationGenerationGaussian SplattingImage
🎯 What it does: In this work, the authors propose PolGS, which combines polarized images and 3D Gaussian Surfels to achieve rapid reconstruction of reflective surfaces within ten minutes.
POMATO: Marrying Pointmap Matching with Temporal Motions for Dynamic 3D Reconstruction
Songyan Zhang (Nanyang Technological University), Chunhua Shen (Zhejiang University)
Object TrackingPose EstimationDepth EstimationTransformerVideoPoint Cloud
🎯 What it does: A unified dynamic 3D reconstruction framework called POMATO has been constructed, capable of simultaneously performing geometric estimation, point cloud matching, and temporal motion modeling, achieving end-to-end 3D reconstruction and point tracking from single frames to videos.
Ponimator: Unfolding Interactive Pose for Versatile Human-human Interaction Animation
Shaowei Liu (University of Illinois Urbana-Champaign), Jian Wang (Snap Inc.)
GenerationPose EstimationDiffusion modelImageVideo
🎯 What it does: Developed the Ponimator framework, which uses two-person close-range interaction poses as anchor points to achieve interactive animation generation.
Pose-Star: Anatomy-Aware Editing for Open-World Fashion Images
Yuran Dong (Wuhan University), Mang Ye (Wuhan University)
Image TranslationSegmentationGenerationDiffusion modelImageMultimodality
🎯 What it does: Proposes the Pose-Star framework, which can dynamically generate fine-grained human masks based on user-specified anatomical prompts (e.g., 'abdominal length') for open-world fashion image editing;
PoseAnchor: Robust Root Position Estimation for 3D Human Pose Estimation
Jun-Hee Kim (Korea University), Seong-Whan Lee (Korea University)
Pose EstimationVideo
🎯 What it does: Proposes the PoseAnchor framework to achieve unsupervised absolute root pose estimation and enhance training robustness through support set selection.
PoseSyn: Synthesizing Diverse 3D Pose Data from In-the-Wild 2D Data
ChangHee Yang (LG Electronics), Hoseok Do (LG Electronics)
GenerationData SynthesisPose EstimationVision Language ModelImageVideoText
🎯 What it does: PoseSyn identifies challenging cases from a large-scale outdoor 2D pose dataset and synthesizes diverse 3D action sequences using text and pose information. It then generates corresponding images through a human animation model, creating a realistic 3D pose training set to enhance the generalization ability of 3D pose estimators.
PossLoss: A Reliable and Sensitive Facial Landmark Detection Loss Function
Qikui Zhu (Case Western Reserve University)
RecognitionPose EstimationConvolutional Neural NetworkImage
🎯 What it does: A position-aware, sample-sensitive heatmap regression loss named PossLoss is proposed to improve facial keypoint detection, addressing peak error, challenging keypoints, and foreground/background imbalance issues.
Power of Cooperative Supervision: Multiple Teachers Framework for Advanced 3D Semi-Supervised Object Detection
Jin-Hee Lee (DGIST), Kwon Soon (DGIST)
Object DetectionAutonomous DrivingPoint Cloud
🎯 What it does: A multiple teacher framework (MultipleTeachers) is proposed for 3D semi-supervised object detection, accompanied by a pseudo-point generator (PointGen) for generating sparse LiDAR point clouds.
PRE-Mamba: A 4D State Space Model for Ultra-High-Frequent Event Camera Deraining
Ciyu Ruan (Shenzhen International Graduate School Tsinghua University), Xinlei Chen (Carnegie Mellon University)
RestorationPoint Cloud
🎯 What it does: This paper proposes a rain algorithm PRE-Mamba based on event cameras, which implements point-level rain denoising using 4D event clouds representation, a spatiotemporal decoupling fusion module (STDF), and a multi-scale state space model (MS3M).
Preacher: Paper-to-Video Agentic System
Jingwei Liu (Peking University), Mengdi Wang (Princeton University)
Large Language ModelAgentic AIVideoTextMultimodalityChain-of-Thought
🎯 What it does: An end-to-end paper-to-video summarization system called Preacher has been developed, which can automatically convert academic papers into structured and visual video summaries.
Precise Action-to-Video Generation Through Visual Action Prompts
Yuang Wang (Zhejiang University), Ruizhen Hu (Shenzhen University)
GenerationData SynthesisRobotic IntelligenceSupervised Fine-TuningPrompt EngineeringVideo
🎯 What it does: Proposes visual action prompts (skeleton rendering) as a unified action representation, constructs a cross-domain data pipeline, and fine-tunes a pre-trained video generation model using ControlNet to achieve precise video generation for complex high-degree-of-freedom interactions.
Predict-Optimize-Distill: A Self-Improving Cycle for 4D Object Understanding
Mingxuan Wu (University of California), Angjoo Kanazawa (University of California)
Pose EstimationOptimizationKnowledge DistillationTransformerGaussian SplattingVideo
🎯 What it does: A self-improving three-stage loop (Prediction-Optimization-Distillation) is proposed to recover the 4D (time + geometry) pose of objects from monocular human-machine interaction videos.
Preserve Anything: Controllable Image Synthesis with Object Preservation
Prasen Kumar Sharma (Typeface), Gaurav Sharma (Typeface)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: The Preserve Anything framework is proposed to achieve high fidelity retention of target objects, semantic consistency, and controllable generation of scene layout and lighting during text-to-image synthesis.
Pretend Benign: A Stealthy Adversarial Attack by Exploiting Vulnerabilities in Cooperative Perception
Hongwei Lin (Xiamen University), Chenglu Wen (Xiamen University)
Object DetectionAutonomous DrivingAdversarial AttackTransformerPoint Cloud
🎯 What it does: A covert adversarial attack named Pretend Benign (PB) is proposed, which exploits the information inconsistency in collaborative perception to disguise the attacker as a trusted collaborator and manipulate the 3D detection results of the target vehicle without being detected.
Pretrained Reversible Generation as Unsupervised Visual Representation Learning
Rongkun Xue (Xi'an Jiaotong University), Jing Yang (Xi'an Jiaotong University)
ClassificationRepresentation LearningDiffusion modelFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: Utilize a pre-trained continuous-time reversible generative model (diffusion/flow model) to extract multi-layer features during the reverse process for downstream unsupervised visual representation learning and classification tasks.
PRIMAL: Physically Reactive and Interactive Motor Model for Avatar Learning
Yan Zhang (Meshcapade), Michael J. Black (Max Planck Institute for Intelligent Systems)
GenerationPose EstimationComputational EfficiencyTransformerDiffusion modelVideo
🎯 What it does: This paper presents PRIMAL, a physics-free autoregressive diffusion model for real-time interactive 3D character animation. The model is first unsupervised pre-trained on large-scale sub-second mocap clips to learn human motion dynamics; it is then efficiently fine-tuned using a ControlNet-style adapter to achieve few-shot personalized action generation and spatial target tracking, and can be rendered in real-time in Unreal Engine.
PrimHOI: Compositional Human-Object Interaction via Reusable Primitives
Kai Jia (Beijing Institute of Technology), Siyuan Huang (National Key Laboratory of General Artificial Intelligence)
Robotic IntelligenceLarge Language ModelDiffusion modelMultimodality
🎯 What it does: The PrimHOI framework is proposed, utilizing reusable interaction primitives (grasping, pinching, supporting, double supporting) to synthesize multi-object human-object interaction actions through hierarchical planning and diffusion models;
Princeton365: A Diverse Dataset with Accurate Camera Pose
Karhan Kayan (Princeton University), Jia Deng (Princeton University)
Pose EstimationOptimizationSimultaneous Localization and MappingOptical FlowVideoMultimodalityBenchmark
🎯 What it does: We propose the Princeton365 dataset, which combines a 360° camera with a calibration board framework to collect 365 segments of multimodal (monocular, stereo, IMU) videos with millimeter-level pose, covering diverse scenes such as indoor, outdoor, and object scanning.
Principles of Visual Tokens for Efficient Video Understanding
Xinyue Hao (University of Edinburgh), Laura Sevilla-Lara (University of Edinburgh)
RecognitionComputational EfficiencyTransformerVideo
🎯 What it does: In the video understanding task, the LITE model is proposed, which learns a lightweight token selector to utilize oracle gradient estimation to assess token value and select a small number of tokens based on importance, significantly reducing computational load;
Prior-aware Dynamic Temporal Modeling Framework for Sequential 3D Hand Pose Estimation
Pengfei Ren (Beijing University of Posts and Telecommunications), Jianxin Liao (Beijing University of Posts and Telecommunications)
Pose EstimationTransformerSequential
🎯 What it does: A prior-aware dynamic temporal modeling framework is proposed for sequential 3D gesture pose estimation.
PriOr-Flow: Enhancing Primitive Panoramic Optical Flow with Orthogonal View
Longliang Liu (Huazhong University of Science and Technology), Xin Yang (Optics Valley Laboratory)
Optical FlowImageVideo
🎯 What it does: A dual-branch framework called PriOr-Flow is proposed, which enhances panoramic optical flow estimation in polar regions using low-distortion priors from orthogonal perspectives.
Prior2Former - Evidential Modeling of Mask Transformers for Assumption-Free Open-World Panoptic Segmentation
Sebastian Schmidt (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
SegmentationAnomaly DetectionTransformerImage
🎯 What it does: This paper proposes Prior2Former (P2F), a vision segmentation model based on MaskTransformer, which utilizes evidence learning to achieve unbiased information estimation of pixel-level masks, completing open-world semantic, panoptic segmentation, and anomaly instance segmentation tasks without relying on any OOD samples.
PriorMotion: Generative Class-Agnostic Motion Prediction with Raster-Vector Motion Field Priors
Kangan Qian (Tsinghua University), Diange Yang (Tsinghua University)
GenerationAutonomous DrivingRecurrent Neural NetworkTransformerAuto EncoderPoint Cloud
🎯 What it does: This paper proposes PriorMotion, a generative framework that models motion priors as distributions in a structured latent space for unconstrained motion prediction.
PRISM: Reducing Spurious Implicit Biases in Vision-Language Models with LLM-Guided Embedding Projection
Mahdiyar Molahasani (Queens University), Ali Etemad (Queens University)
ClassificationRecognitionOptimizationTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: PRISM is proposed, a data-free, task-agnostic visual-language model debiasing method that utilizes LLMs to generate potential implicit bias scene descriptions and learns projections to remove spurious correlations in the embedding space.
Privacy-centric Deep Motion Retargeting for Anonymization of Skeleton-Based Motion Visualization
Thomas Carr (University of North Carolina at Charlotte), Aidong Lu (University of North Carolina at Charlotte)
RecognitionSafty and PrivacyAuto EncoderGenerative Adversarial NetworkVideo
🎯 What it does: A privacy protection framework based on deep motion redirection (PMR) is proposed, which achieves skeleton data anonymization by transferring user movements to a false skeleton.
PRM: Photometric Stereo based Large Reconstruction Model
Wenhang Ge (Hong Kong University of Science and Technology GZ), Ying-Cong Chen (Hong Kong University of Science and Technology)
RestorationGenerationTransformerNeural Radiance FieldMesh
🎯 What it does: A Large Reconstruction Model (PRM) based on photometric stereo has been developed, which can directly generate high-quality, detail-rich meshes and textures by using images with variable materials and lighting during training, along with differentiable PBR rendering.
PRO-VPT: Distribution-Adaptive Visual Prompt Tuning via Prompt Relocation
Chikai Shang (Xiamen University), Yiu-Ming Cheung (Hong Kong Baptist University)
ClassificationObject DetectionSegmentationOptimizationTransformerReinforcement LearningPrompt EngineeringImage
🎯 What it does: The PRO-VPT framework is proposed, which dynamically adjusts the distribution of visual prompts through an iterative Prompt Relocation (PR) mechanism, achieving efficient parameterized fine-tuning of pre-trained visual models.
Proactive Scene Decomposition and Reconstruction
Baicheng Li (Peking University), Hongbin Zha (Peking University)
Object DetectionSegmentationPose EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingVideo
🎯 What it does: This paper proposes the task of Proactive Scene Decomposition and Reconstruction, which utilizes hand-object interaction information from an egocentric perspective to incrementally decompose dynamic environments into independently moving objects and perform high-quality reconstruction.
Probabilistic Inertial Poser (ProbIP): Uncertainty-aware Human Motion Modeling from Sparse Inertial Sensors
Min Kim (Korea Advanced Institute of Science and Technology), Sungho Jo (Korea Advanced Institute of Science and Technology)
Pose EstimationGenerative Adversarial NetworkTime Series
🎯 What it does: Proposes the ProbIP model, which utilizes sparse IMU sensors to reconstruct full-body posture in a probabilistic manner, achieving real-time motion prediction without physical constraints.
Probabilistic Prototype Calibration of Vision-language Models for Generalized Few-shot Semantic Segmentation
Jie Liu (University of Amsterdam), Efstratios Gavves
SegmentationTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: A FewCLIP framework based on probabilistic prototype calibration is proposed for generalizing few-shot semantic segmentation, which can better capture intra-class variations and enhance generalization ability when sample scarcity occurs in new categories.
ProbMED: A Probabilistic Framework for Medical Multimodal Binding
Yuan Gao (University of Toronto), Chris McIntosh (University of Toronto)
ClassificationRetrievalTransformerContrastive LearningMultimodalityBiomedical DataUltrasoundElectrocardiogram
🎯 What it does: Proposes ProbMED, a multimodal visual-language pre-training framework that aligns lung X-rays, electrocardiograms, echocardiograms, and clinical texts through probabilistic alignment.
ProbRes: Probabilistic Jump Diffusion for Open-World Egocentric Activity Recognition
Sanjoy Kundu (Auburn University), Sathyanarayanan N. Aakur (Auburn University)
RecognitionOptimizationComputational EfficiencyTransformerVision Language ModelDiffusion modelVideo
🎯 What it does: This paper proposes ProbRes, a probabilistic residual search framework based on jump diffusion, designed for efficiently exploring vast search spaces in open-world first-person activity recognition.
Processing and acquisition traces in visual encoders: What does CLIP know about your camera?
Ryan Ramos (Osaka University), Noa Garcia (Osaka University)
RetrievalTransformerVision Language ModelContrastive LearningImage
🎯 What it does: Analyze whether the visual encoder retains traces of image processing and acquisition parameters in the embeddings, and assess their impact on semantic tasks.
ProGait: A Multi-Purpose Video Dataset and Benchmark for Transfemoral Prosthesis Users
Xiangyu Yin (University of Pittsburgh), Wei Gao (University of Pittsburgh)
ClassificationObject DetectionSegmentationPose EstimationRecurrent Neural NetworkSupervised Fine-TuningVideoTextBenchmark
🎯 What it does: This paper presents a multi-task video dataset named ProGait, specifically designed for video object segmentation, 2D human pose estimation, and gait analysis tasks related to amputees.
Progressive Artwork Outpainting via Latent Diffusion Models
Dae-Young Song (Electronics and Telecommunications Research Institute), Donghyeon Cho (Hanyang University)
GenerationDiffusion modelImage
🎯 What it does: This paper studies an evolvable method for the extension drawing of artistic works, combining a stepwise expansion approach with a pre-trained stable diffusion model to achieve high-resolution works with arbitrary aspect ratios.
Progressive Distribution Bridging: Unsupervised Adaptation for Large-scale Pre-trained Models via Adaptive Auxiliary Data
Weinan He (University of Science and Technology of China), Zilei Wang (University of Science and Technology of China)
Domain AdaptationVision Language ModelContrastive LearningMultimodality
🎯 What it does: An unsupervised adaptive method is proposed, which helps large-scale pre-trained visual language models adapt to unlabeled target domains by retrieving and gradually adjusting the auxiliary data distribution from large-scale image-text data.
Progressive Growing of Video Tokenizers for Temporally Compact Latent Spaces
Aniruddha Mahapatra (Adobe Research), Feng Liu (Adobe Research)
GenerationCompressionConvolutional Neural NetworkGenerative Adversarial NetworkVideo
🎯 What it does: A video tokenizer named ProMAG is proposed, which achieves efficient video latent representation by progressively increasing the time compression ratio from 4× to 8× and 16× while keeping the channel dimension unchanged.
Progressive Homeostatic and Plastic Prompt Tuning for Audio-Visual Multi-Task Incremental Learning
Jiong Yin (Hangzhou Dianzi University), Xichun Sheng (Macao Polytechnic University)
RecognitionSegmentationData-Centric LearningTransformerPrompt EngineeringContrastive LearningVideoMultimodalityAudio
🎯 What it does: A three-stage evolutionary framework for prompt-based stability and plasticity (PHP) is proposed to address the issues of catastrophic forgetting and task interference in audio-visual multi-task incremental learning. The framework introduces a task-shared modality aggregation adapter (TMA) at the shallow level, uses a task-specific but modality-shared dynamic generation adapter (TMDG) at the middle level, and adds task-specific and modality-independent prompts (TMI) at the deep level, thereby achieving a layered knowledge transfer from commonality to personalization.
Progressive Test Time Energy Adaptation for Medical Image Segmentation
Xiaoran Zhang (Yale University), Alex Wong (Yale University)
SegmentationDomain AdaptationAdversarial AttackConvolutional Neural NetworkBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A stepwise testing energy adaptive method for medical image segmentation is proposed, which dynamically updates the pre-trained model during testing to adapt to the target distribution.
PROGRESSOR: A Perceptually Guided Reward Estimator with Self-Supervised Online Refinement
Tewodros W. Ayalew (University of Chicago), Matthew R. Walter (Toyota Technological Institute at Chicago)
Robotic IntelligenceReinforcement LearningVideo
🎯 What it does: Proposes the PROGRESSOR framework, which uses a task-agnostic reward function for unsupervised learning from unlabeled videos to guide reinforcement learning through progress prediction.
ProJudge: A Multi-Modal Multi-Discipline Benchmark and Instruction-Tuning Dataset for MLLM-based Process Judges
Jiaxin Ai (Wuhan University), Kaipeng Zhang (Shanghai AI Lab)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmarkPhysics Related
🎯 What it does: This paper proposes a benchmark called ProJudgeBench specifically designed to evaluate the reasoning process of multimodal large language models (MLLMs) and constructs a large-scale instruction tuning dataset, ProJudge-173k. It also introduces a Dynamic Dual-Phase fine-tuning strategy to enhance the process evaluation capability of open-source models.
PROL : Rehearsal Free Continual Learning in Streaming Data via Prompt Online Learning
M. Anwar Ma'sum (University of South Australia), Ryszard Kowalczyk (Polish Academy of Sciences)
ClassificationRecognitionTransformerPrompt EngineeringImage
🎯 What it does: This paper proposes a replay-free online continual learning framework called PROL, which utilizes a single lightweight prompt generator and learnable scale-offset parameters to achieve stable-plastic learning on pre-trained models.
Prompt Guidance and Human Proximal Perception for HOT Prediction with Regional Joint Loss
Yuxiao Wang (South China University of Technology), Qi Liu (South China University of Technology)
Object DetectionSegmentationConvolutional Neural NetworkPrompt EngineeringImage
🎯 What it does: A HOT (Human-Object Contact) detection framework called P3HOT is proposed, which combines text prompts and human proximal perception to significantly enhance the localization and classification of contact areas using depth information and textual guidance.
Prompt-A-Video: Prompt Your Video Diffusion Model via Preference-Aligned LLM
Yatai Ji (University of Hong Kong), Ping Luo (University of Hong Kong)
GenerationData SynthesisOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringDiffusion modelVideoText
🎯 What it does: This paper proposes and implements Prompt-A-Video, an automatic video prompt optimization system based on large language models, capable of generating high-quality video prompts that align with the preferences of video generation models.
Prompt-driven Transferable Adversarial Attack on Person Re-Identification with Attribute-aware Textual Inversion
Yuan Bian (Hunan University), Yaonan Wang (Hunan University)
RecognitionAdversarial AttackTransformerVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes a transferable adversarial attack based on attribute-aware prompts (AP-Attack). By leveraging the image-text alignment capability of the visual-language model (CLIP), pedestrian images are mapped to attribute-level pseudo-words. These attribute semantics are utilized in the adversarial generator for fine-grained semantic disruption, resulting in adversarial samples with a high attack success rate across models and datasets.
PromptDresser: Improving the Quality and Controllability of Virtual Try-On via Generative Textual Prompt and Prompt-aware Mask
Jeongho Kim (KAIST), Jaegul Choo (KAIST)
GenerationData SynthesisLarge Language ModelPrompt EngineeringDiffusion modelImageText
🎯 What it does: The PromptDresser model is proposed to achieve editable virtual try-ons through generative text prompts and adaptive masks for high-quality and controllable clothing synthesis.
PropVG: End-to-End Proposal-Driven Visual Grounding with Multi-Granularity Discrimination
Ming Dai (Southeast University), Wankou Yang (Southeast University)
Object DetectionSegmentationTransformerContrastive LearningImageText
🎯 What it does: An end-to-end proposal-driven visual localization framework, PropVG, is proposed, which integrates foreground proposal generation, Contrastive Reference Scoring (CRS), and Multi-Granularity Target Discrimination (MTD), achieving visual localization and segmentation without the need for a pre-trained detector.
ProSAM: Enhancing the Robustness of SAM-based Visual Reference Segmentation with Probabilistic Prompts
Xiaoqi Wang (Bosch Research North America), Liu Ren (Bosch Research North America)
SegmentationDomain AdaptationPrompt EngineeringAuto EncoderImage
🎯 What it does: This paper proposes ProSAM, a probabilistic prompt generation method based on variational inference, aimed at enhancing the robustness and zero-shot performance of SAM in visual reference segmentation.
Prototype Guided Backdoor Defense via Activation Space Manipulation
Venkat Adithya Amula (International Institute of Information Technology Hyderabad), P J Narayanan (International Institute of Information Technology Hyderabad)
ClassificationAdversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a post-hoc defense method against backdoor attacks, utilizing Prototype Activation Vectors (PAV) to impose geometric constraints on the model's activation space, thereby eliminating misclassifications caused by malicious triggers.
Prototype-based Contrastive Learning with Stage-wise Progressive Augmentation for Self-Supervised Fine-Grained Learning
Baofeng Tan (Nanjing University of Science and Technology), Lin Zhao (Nanjing University of Science and Technology)
ClassificationRetrievalRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: The PAPN method is proposed, which improves self-supervised fine-grained visual representation through stage-wise evolution enhancement and prototype contrastive learning.
Prototypes are Balanced Units for Efficient and Effective Partially Relevant Video Retrieval
WonJun Moon (NAVER Cloud), Jae-Pil Heo (NAVER Cloud)
RetrievalTransformerVision Language ModelVideoText
🎯 What it does: A prototype-based Partial Relevant Video Retrieval (PRVR) framework is proposed, which utilizes a small number of prototypes to aggregate the context of videos at different time scales, and enhances text relevance and video understanding through dual-modal reconstruction, weak supervision guidance, and orthogonal regularization.
Proxy-Bridged Game Transformer for Interactive Extreme Motion Prediction
Yanwen Fang (University of Hong Kong), Jintai Chen
Pose EstimationTransformerTime SeriesSequential
🎯 What it does: A Transformer-based multi-person extreme interaction motion prediction framework called PGformer is proposed, which can predict future 3D poses from past pose sequences.
Pruning All-Rounder: Rethinking and Improving Inference Efficiency for Large Vision Language Models
Wei Suo (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
OptimizationComputational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: A framework named Pruning All-Rounder (PAR) is proposed, which can significantly reduce inference FLOPs by jointly pruning visual tokens and model layers through a self-supervised preference optimization method without retraining the weights of large visual language models.
PRVQL: Progressive Knowledge-guided Refinement for Robust Egocentric Visual Query Localization
Bing Fan (University of North Texas), Heng Fan (University of Texas at Dallas)
Object DetectionRetrievalTransformerVideo
🎯 What it does: A novel evolutionary knowledge-guided refinement framework named PRVQL is proposed to enhance the accuracy of locating visual query objects from first-person perspective videos.
PS-Mamba: Spatial-Temporal Graph Mamba for Pose Sequence Refinement
Haoye Dong (National University of Singapore), Gim Hee Lee (National University of Singapore)
Pose EstimationGraph Neural NetworkSequential
🎯 What it does: Refine the human posture sequence to make it smoother over time and more coherent in structure.
PS3: A Multimodal Transformer Integrating Pathology Reports with Histology Images and Biological Pathways for Cancer Survival Prediction
Manahil Raza (University of Warwick), Nasir Rajpoot (University of Warwick)
ClassificationTransformerVision Language ModelImageTextMultimodalityBiomedical Data
🎯 What it does: Developed the PS3 model, which utilizes prototype representations of pathology reports, whole slide images, and transcriptome data for multimodal fusion to predict cancer survival time.
Pseudo-SD: Pseudo Controlled Stable Diffusion for Semi-Supervised and Cross-Domain Semantic Segmentation
Dong Zhao (University of Trento), Zhun Zhong (Hefei University of Technology)
SegmentationDomain AdaptationDiffusion modelImage
🎯 What it does: This paper proposes the Pseudo-SD framework, which utilizes pseudo-labels to generate high-quality and diverse synthetic semantic segmentation data through Stable Diffusion for semi-supervised and cross-domain semantic segmentation tasks.
PseudoMapTrainer: Learning Online Mapping without HD Maps
Christian Löwens (Bosch), Alexandru Paul Condurache (Bosch)
SegmentationAutonomous DrivingGaussian SplattingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: A method is proposed to train an online mapping model without HD map labels, and performance is improved through pseudo-label pre-training.
PUMA: Empowering Unified MLLM with Multi-granular Visual Generation
Rongyao Fang (Chinese University of Hong Kong), Xihui Liu (Hong Kong University)
Image TranslationGenerationTransformerLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: A unified multimodal large language model PUMA is proposed, capable of simultaneously performing various visual generation and understanding tasks through multi-granularity visual features, including text-to-image, diverse generation, editing, inpainting, coloring, and conditional generation.
PUMPS: Skeleton-Agnostic Point-based Universal Motion Pre-Training for Synthesis in Human Motion Tasks
Clinton Ansun Mo (University of Sydney), Zhiyong Wang (University of Sydney)
GenerationData SynthesisPose EstimationTransformerAuto EncoderVideoPoint Cloud
🎯 What it does: A skeleton-independent temporal point cloud autoencoder PUMPS is proposed for general motion pre-training and synthesis;
Punching Bag vs. Punching Person: Motion Transferability in Videos
Raiyaan Abdullah (University of Central Florida), Yogesh Rawat (University of Central Florida)
ClassificationRecognitionDomain AdaptationConvolutional Neural NetworkContrastive LearningVideoMultimodalityBenchmark
🎯 What it does: This paper studies the motion transferability of action recognition models in different contexts, proposing three benchmark datasets (Syn-TA, Kinetics400-TA, Something‑Something‑v2‑TA) and systematically evaluating 13 state-of-the-art models.
Puppet-Master: Scaling Interactive Video Generation as a Motion Prior for Part-Level Dynamics
Ruining Li (Visual Geometry Group, University of Oxford), Andrea Vedaldi (Visual Geometry Group, University of Oxford)
GenerationData SynthesisDiffusion modelVideo
🎯 What it does: We propose Puppet-Master, an interactive video generation model based on diffusion, which allows control of the motion of internal components of objects through a small number of drag points.
Purge-Gate: Backpropagation-Free Test-Time Adaptation for Point Clouds Classification via Token purging
Moslem Yazdanpanah (École supérieure de technologie de Montréal), Christian Desrosiers (École supérieure de technologie de Montréal)
ClassificationDomain AdaptationTransformerPoint Cloud
🎯 What it does: This paper proposes a token purging-based method for test-time adaptation (TTA) in point cloud classification, called Purge-Gate (PG), which enhances robustness by removing the tokens most affected by domain shift before the attention layer.
Puzzle Similarity: A Perceptually-guided Cross-Reference Metric for Artifact Detection in 3D Scene Reconstructions
Nicolai Hermann (Università della Svizzera italiana), Piotr Didyk (Istituto Dalle Molle di Studi sull'Intelligenza Artificiale)
RestorationAnomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: A no-reference, cross-view image quality metric called Puzzle Similarity is designed to detect and locate artifacts in newly reconstructed 3D views.
PVChat: Personalized Video Chat with One-Shot Learning
Yufei Shi (Nanyang Technological University), Si Yong Yeo (Nanyang Technological University)
GenerationData SynthesisRecommendation SystemTransformerLarge Language ModelVision Language ModelVideoText
🎯 What it does: PVChat is proposed, a ViLLM framework capable of personalized video question answering from a single video;
PVMamba: Parallelizing Vision Mamba via Dynamic State Aggregation
Fei Xie (MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University), Chao Ma (MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerImage
🎯 What it does: A visual Mamba named PVMamba is proposed, achieving parallel computation through Dynamic State Aggregation (DSA) to address the sequential constraints of traditional Mamba.
Q-Frame: Query-aware Frame Selection and Multi-Resolution Adaptation for Video-LLMs
Shaojie Zhang (Xiaomi Inc), Jian Luan (Xiaomi Inc)
RetrievalComputational EfficiencyTransformerVision Language ModelVideo
🎯 What it does: This paper addresses the frame sampling bottleneck of video large language models and proposes the Q-Frame framework, which can dynamically select key signal frames based on queries and adaptively adjust resolution according to importance.
Q-Norm: Robust Representation Learning via Quality-Adaptive Normalization
Lanning Zhang (Xidian University), Nannan Wang (Xidian University)
Image TranslationObject DetectionDomain AdaptationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: A quality-adaptive normalization method called Q-Norm is proposed to enhance the robustness of deep networks on low-quality images.
QK-Edit: Revisiting Attention-based Injection in MM-DiT for Image and Video Editing
Tiancheng Shen (Chinese University of Hong Kong), Jun Hao Liew (ByteDance Seed)
GenerationData SynthesisTransformerDiffusion modelImageVideo
🎯 What it does: A training-free text-guided editing framework called QK-Edit is proposed, which utilizes the reconstruction of queries and keys in the Multimodal Diffusion Transformer (MM-DiT) to achieve high-fidelity image and video editing.
QR-LoRA: Efficient and Disentangled Fine-tuning via QR Decomposition for Customized Generation
Jiahui Yang (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes QR-LoRA, a method that utilizes QR decomposition for parameter-efficient and decoupled fine-tuning, allowing for independent control of content and style features in text-to-image models.
Quadratic Gaussian Splatting: High Quality Surface Reconstruction with Second-order Geometric Primitives
Ziyu Zhang (Chinese Academy of Sciences Institute of Automation), Shuhan Shen (SenseTime Research)
RestorationOptimizationGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes Quadratic Gaussian Splatting (QGS), which uses deformable quadratic surfaces to replace traditional planar or spherical primitives for high-quality surface reconstruction.
Quanta Neural Networks: From Photons to Perception
Varun Sundar (University of Wisconsin-Madison), Mohit Gupta (University of Wisconsin-Madison)
RestorationObject TrackingDepth EstimationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageVideo
🎯 What it does: A quantum neural network (QNN) is proposed to directly infer photon sequences captured by quantum cameras, replacing convolution/transformer layers in traditional image/video networks with learnable temporal QNN layers, achieving end-to-end perception without image reconstruction.
QuantCache: Adaptive Importance-Guided Quantization with Hierarchical Latent and Layer Caching for Video Generation
Junyi Wu (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
GenerationOptimizationComputational EfficiencyTransformerDiffusion modelVideo
🎯 What it does: This paper presents QuantCache, a training-independent video generation acceleration framework that jointly optimizes hierarchical latent caching, importance-based adaptive quantization, and structural redundancy-aware pruning, significantly improving the inference speed of Diffusion Transformers.
Quantifying and Narrowing the Unknown: Interactive Text-to-Video Retrieval via Uncertainty Minimization
Bingqing Zhang (University of Queensland), Sen Wang (University of Queensland)
RetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoText
🎯 What it does: An interactive text-to-video retrieval framework UMIVR is proposed, which explicitly quantifies and minimizes three types of uncertainties: text ambiguity, mapping uncertainty, and frame quality uncertainty, and iteratively improves queries through adaptive clarification questions.
QuEST: Low-bit Diffusion Model Quantization via Efficient Selective Finetuning
Haoxuan Wang (University of Illinois Chicago), Yan Yan (University of Illinois Chicago)
GenerationData SynthesisOptimizationDiffusion modelImage
🎯 What it does: Proposes the QuEST framework, which achieves low-bit diffusion model quantization through efficient selective fine-tuning.
QuickSplat: Fast 3D Surface Reconstruction via Learned Gaussian Initialization
Yueh-Cheng Liu (Technical University of Munich), Angela Dai (Technical University of Munich)
Depth EstimationOptimizationConvolutional Neural NetworkNeural Radiance FieldGaussian SplattingPoint CloudMesh
🎯 What it does: Utilizing data-driven priors for rapid 3D surface reconstruction of multi-view indoor scenes, employing 2D Gaussian Splatting and using learned initialization, densification, and update networks in iterative optimization, significantly enhances convergence speed and reconstruction accuracy.
R-LiViT: A LiDAR-Visual-Thermal Dataset Enabling Vulnerable Road User Focused Roadside Perception
Jonas Mirlach (XITASO GmbH), Andreas Eich (LiangDao GmbH)
Object DetectionObject TrackingAutonomous DrivingImageMultimodalityPoint Cloud
🎯 What it does: A three-modal dataset R-LiViT of LiDAR, RGB, and thermal imaging for roadside intersections has been constructed, with a focus on annotating VRUs.
R1-Onevision: Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization
Yi Yang (Zhejiang University), Wei Chen (Zhejiang University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper presents R1-Onevision, a multimodal reasoning framework based on cross-modal formalization, which includes the conversion of images to formal text, role-playing reasoning generation, two-stage post-training (SFT+RL), and evaluation benchmarks dimensioned by educational grade.
R1-VL: Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization
Jingyi Zhang (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
TransformerLarge Language ModelReinforcement LearningVision Language ModelTextMultimodality
🎯 What it does: This paper proposes a new online reinforcement learning framework called StepGRPO, which enhances the reasoning capabilities of multimodal large language models (MLLMs) using step-level rewards.