ICCV 2025 Papers — Page 18
IEEE/CVF International Conference on Computer Vision · 2701 papers
OmniSAM: Omnidirectional Segment Anything Model for UDA in Panoramic Semantic Segmentation
Ding Zhong (Hong Kong University of Science and Technology), Xuming Hu (Hong Kong University of Science and Technology)
SegmentationDomain AdaptationImage
🎯 What it does: The OmniSAM framework is proposed, which migrates SAM2 to the unsupervised domain adaptation task of 360° panoramic semantic segmentation. It achieves cross-domain feature alignment by splitting images using a sliding window, employing a memory mechanism, and utilizing a semantic decoder.
OmniVTON: Training-Free Universal Virtual Try-On
Zhaotong Yang (Ocean University of China), Yong Du (Ocean University of China)
Image TranslationGenerationPose EstimationDiffusion modelImage
🎯 What it does: A training-free universal virtual try-on framework called OmniVTON is proposed, which can seamlessly synthesize clothing images onto target portraits in both indoor and outdoor scenes without training, and supports multi-person try-on.
On Large Multimodal Models as Open-World Image Classifiers
Alessandro Conti (University of Trento), Elisa Ricci (Fondazione Bruno Kessler)
ClassificationTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: Evaluate the image classification capability of large-scale multimodal models (LMM) under unrestricted categories (open world) and propose four comprehensive evaluation metrics.
On the Complexity-Faithfulness Trade-off of Gradient-Based Explanations
Amir Mehrpanah (KTH Royal Institute of Technology), Hossein Azizpour (KTH Royal Institute of Technology)
Explainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: A unified spectral framework is proposed to quantify the trade-off between the complexity and fidelity of gradient explanations, and the complexity of explanations is reduced by controlling the high-frequency information of ReLU.
On the Generalization of Representation Uncertainty in Earth Observation
Spyros Kondylatos (National Observatory of Athens), Ioannis Papoutsis (National Technical University of Athens)
ClassificationSegmentationTransformerImage
🎯 What it does: In this paper, the authors pre-trained a representation uncertainty module on a large-scale Earth Observation (EO) dataset and conducted zero-shot evaluation on downstream EO tasks such as multi-label classification and semantic segmentation, proposing an uncertainty assessment framework and metrics suitable for multi-label and segmentation tasks.
On the Provable Importance of Gradients for Autonomous Language-Assisted Image Clustering
Bo Peng (University of Technology Sydney), Zhen Fang (University of Technology Sydney)
ClassificationRecognitionContrastive LearningImageText
🎯 What it does: This paper proposes GradNorm, a gradient-based framework for selecting positive nouns from unlabeled text corpora to assist image clustering;
On the Recovery of Cameras from Fundamental Matrices
Rakshith Madhavan (Politecnico di Milano), Federica Arrigoni (Politecnico di Milano)
OptimizationImage
🎯 What it does: A method is proposed for recovering a camera from a noisy Fundamental matrix for any solvable view graph without calibration.
On the Robustness Tradeoff in Fine-Tuning
Kunyang Li (University of Wisconsin Madison), Patrick McDaniel (University of Wisconsin Madison)
ClassificationDomain AdaptationAdversarial AttackTransformerSupervised Fine-TuningImage
🎯 What it does: This paper systematically evaluates the trade-off between robustness and accuracy of pre-trained models in downstream tasks under different fine-tuning strategies.
On-Device Diffusion Transformer Policy for Efficient Robot Manipulation
Yiming Wu (University of Hong Kong), Dong Xu (University of Hong Kong)
Computational EfficiencyKnowledge DistillationRobotic IntelligenceTransformerDiffusion modelMultimodality
🎯 What it does: The LightDP framework is proposed to achieve real-time robot control on mobile devices by compressing diffusion strategies.
One Encoder to Rule them All: Representation Learning for Model-free Visual Reinforcement Learning using Fourier Neural Operators
Parag Dutta (Indian Institute of Science), Ambedkar Dukkipati (Indian Institute of Science)
Autonomous DrivingRepresentation LearningConvolutional Neural NetworkReinforcement LearningContrastive LearningImage
🎯 What it does: This paper proposes replacing the traditional CNN image encoder with a Fourier Neural Operator (FNO) encoder to learn the underlying PDE dynamics of the environment in model-free visual reinforcement learning and to directly extract features from images.
One Last Attention for Your Vision-Language Model
Liang Chen (Mohammed Bin Zayed University of Artificial Intelligence), Zhiqiang Shen (Mohammed Bin Zayed University of Artificial Intelligence)
ClassificationDomain AdaptationComputational EfficiencyTransformerSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: A lightweight attention adapter (RAda) based on Rational Matrix is proposed, which directly adjusts the final fusion representation of VLM through adaptive masking, achieving efficient fine-tuning of the multimodal decision process.
One Look is Enough: Seamless Patchwise Refinement for Zero-Shot Monocular Depth Estimation on High-Resolution Images
Byeongjun Kwon (Korea Advanced Institute of Science and Technology), Munchurl Kim (Korea Advanced Institute of Science and Technology)
Depth EstimationImage
🎯 What it does: Proposes the Patch Refine Once (PRO) framework, which achieves zero-shot inference for high-resolution monocular depth estimation through single patch refinement.
One Object, Multiple Lies: A Benchmark for Cross-task Adversarial Attack on Unified Vision-Language Models
Jiale Zhao (Tongji University), Cairong Zhao (Tongji University)
Adversarial AttackVision Language ModelContrastive LearningImageTextBenchmark
🎯 What it does: This paper proposes the CrossVLAD benchmark for cross-task adversarial attack evaluation and designs the CRAFT framework for region-token alignment to achieve a single perturbation that deceives multiple downstream tasks on a unified vision-language model.
One Perturbation is Enough: On Generating Universal Adversarial Perturbations against Vision-Language Pre-training Models
Hao Fang (Tsinghua University), Ke Xu (Tsinghua University)
RetrievalAdversarial AttackTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a general adversarial perturbation generation framework C-PGC based on contrastive learning, which generates adversarial perturbations for images and texts in one go to deceive visual-language pre-trained models.
One Polyp Identifies All: One-Shot Polyp Segmentation with SAM via Cascaded Priors and Iterative Prompt Evolution
Xinyu Mao (Chinese University of Hong Kong), Max Meng (Southern University of Science and Technology)
SegmentationTransformerPrompt EngineeringImageBiomedical Data
🎯 What it does: A single-sample polyp segmentation framework based on SAM, OP-SAM, is proposed, which automatically generates prompt points from a single annotated image, achieving zero training and extremely low annotation costs.
One Trajectory, One Token: Grounded Video Tokenization via Panoptic Sub-object Trajectory
Chenhao Zheng (University of Washington), Ranjay Krishna (University of Washington)
Object TrackingSegmentationRetrievalCompressionTransformerVision Language ModelContrastive LearningVideoText
🎯 What it does: This paper proposes 'Grounded Video Tokenization' based on panoptic sub-object trajectories, achieving efficient video encoding by decomposing videos into object trajectories and encoding each trajectory as a single token.
One-Shot Knowledge Transfer for Scalable Person Re-Identification
Longhua Li (Southeast University), Xin Geng (Southeast University)
RecognitionRetrievalCompressionKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: A one-click knowledge transfer framework OSKT is proposed, which constructs a weight chain on the teacher model, allowing for the generation of person re-identification models of any size with a single computation.
One-Step Specular Highlight Removal with Adapted Diffusion Models
Mahir Atmis (Cukurova University), Mehmet Sarıgül
RestorationDiffusion modelImage
🎯 What it does: This paper proposes a single-step diffusion model based on Stable Diffusion Turbo, utilizing the ProbLoRA adapter to achieve high light removal;
OneGT: One-Shot Geometry-Texture Neural Rendering for Head Avatars
Jinshu Chen (Bytedance Intelligent Creation), Qian He (Bytedance Intelligent Creation)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImage
🎯 What it does: We propose OneGT, a neural rendering framework for geometry-texture decoupling from a single sample, achieving editable and high-fidelity head avatars.
Online Dense Point Tracking with Streaming Memory
Qiaole Dong (Fudan University), Yanwei Fu (Fudan University)
Object TrackingRecurrent Neural NetworkOptical FlowVideo
🎯 What it does: We propose SPOT, a framework that utilizes streaming memory to achieve efficient online dense point tracking, capable of estimating dense optical flow and point trajectories without relying on future frames.
Online Generic Event Boundary Detection
Hyungrok Jung (Gwangju Institute of Science and Technology), Jonghyun Choi (Seoul National University)
Anomaly DetectionTransformerVideo
🎯 What it does: This paper proposes the Online General Event Boundary Detection task (On-GEBD) and presents a detection framework called ESTimator based on event segmentation theory.
Online Language Splatting
Saimouli Katragadda (University of Delaware), Liu Ren (Bosch Research North America)
Super ResolutionOptimizationComputational EfficiencyAuto EncoderGaussian SplattingSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: A real-time, online, open-vocabulary 3D language mapping framework has been developed, integrating CLIP language features into 3D Gaussian distribution SLAM.
Online Reasoning Video Segmentation with Just-in-Time Digital Twins
Yiqing Shen (Johns Hopkins University), Mathias Unberath (Johns Hopkins University)
Object DetectionSegmentationTransformerLarge Language ModelAgentic AIVideoBenchmark
🎯 What it does: A proxy-based framework is proposed, utilizing real-time digital twins to achieve online video inference segmentation, supporting multi-step semantic, spatial, and temporal reasoning.
ONLY: One-Layer Intervention Sufficiently Mitigates Hallucinations in Large Vision-Language Models
Zifu Wan (Carnegie Mellon University), Yaqi Xie (Carnegie Mellon University)
RecognitionGenerationComputational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: This paper proposes ONLY, a training method that requires only a single-layer intervention and no additional queries, aimed at significantly reducing the hallucination problem in large visual language models.
Open-ended Hierarchical Streaming Video Understanding with Vision Language Models
Hyolim Kang (Yonsei University), Seon Joo Kim (Yonsei University)
RecognitionSegmentationRetrievalRecurrent Neural NetworkTransformerLarge Language ModelVision Language ModelVideoText
🎯 What it does: Proposes the Hierarchical Streaming Video Understanding task, combining real-time action localization with free-text descriptions; introduces the OpenHOUSE framework, achieving a separation of lightweight streaming perception modules and frozen VLM calls to complete instant hierarchical descriptions.
Open-set Cross Modal Generalization via Multimodal Unified Representation
Hai Huang (Zhejiang University), Zhou Zhao (Zhejiang University)
RecognitionRetrievalDomain AdaptationContrastive LearningVideoMultimodalityAudio
🎯 What it does: This paper proposes the Open Set Cross-Modal Generalization (OSCMG) task and designs the MICU framework to achieve unified representation while enhancing recognition performance for unknown categories.
Open-Unfairness Adversarial Mitigation for Generalized Deepfake Detection
Zhaoyang Li (Beijing Jiaotong University), Jianping Fan (Lenovo Research)
ClassificationAnomaly DetectionGenerative Adversarial NetworkVideo
🎯 What it does: Proposed the Adversarial Open Unfairness Discovery and Mitigation Network (AdvOU), which embeds a lightweight Unfairness Regulator in the deepfake detection model and alternately trains the OUD and UAM modules to achieve dynamic identification and elimination of undefined biases.
Open-Vocabulary HOI Detection with Interaction-aware Prompt and Concept Calibration
Ting Lei (Wangxuan Institute of Computer Technology, Peking University), Yang Liu (Wangxuan Institute of Computer Technology, Peking University)
RecognitionObject DetectionTransformerLarge Language ModelPrompt EngineeringContrastive LearningMultimodality
🎯 What it does: This paper proposes an open vocabulary human-object interaction (HOI) detection framework called INP-CC, which utilizes Interaction-Aware Prompts and Concept Calibration to address the distribution bias and semantic alignment issues in traditional CLIP for region-level interaction recognition.
Open-Vocabulary Octree-Graph for 3D Scene Understanding
Zhigang Wang (Northwestern Polytechnical University), Bin Zhao (University of Chinese Academy of Sciences)
SegmentationRetrievalGraph Neural NetworkLarge Language ModelVision Language ModelPoint Cloud
🎯 What it does: Proposes Octree-Graph as a new scene representation for open vocabulary 3D scene understanding, combining adaptive octree nodes with graph structures to achieve efficient occupancy, semantic, and spatial relationship modeling.
Open-World Skill Discovery from Unsegmented Demonstration Videos
Jingwen Deng (Peking University), Yitao Liang (Peking University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningAgentic AIVideo
🎯 What it does: This paper proposes an unsupervised Skill Boundary Detection (SBD) method that utilizes a pre-trained unconditional action prediction model to detect skill transition points in long, unsegmented videos, thereby automatically segmenting lengthy demonstration videos into independent skill segments. Based on this, it trains video/text conditional policies and hierarchical agents.
OpenAnimals: Revisiting Person Re-Identification for Animals Towards Better Generalization
Saihui Hou (Beijing Normal University), Yongzhen Huang (Beijing Normal University)
RecognitionConvolutional Neural NetworkSupervised Fine-TuningImageBenchmark
🎯 What it does: An open-source code library for animal re-identification, OpenAnimals, has been constructed, and based on it, the applicability of various human re-identification methods to animal data has been re-evaluated, ultimately proposing a benchmark model specifically for animal re-identification, ARBase.
OpenM3D: Open Vocabulary Multi-view Indoor 3D Object Detection without Human Annotations
Peng-Hao Hsu (National Tsing Hua University), Cheng-Hao Kuo (National Tsing Hua University)
Object DetectionSegmentationContrastive LearningImage
🎯 What it does: A single-stage indoor 3D object detection model called OpenM3D is proposed, which is based on multi-view RGB images and does not require manual 3D box annotations, supporting open vocabulary classification.
OpenRSD: Towards Open-prompts for Object Detection in Remote Sensing Images
Ziyue Huang (Beihang University), Yunhong Wang (Beihang University)
Object DetectionContrastive LearningImageMultimodality
🎯 What it does: A multi-modal prompt-based remote sensing object detection framework called OpenRSD is proposed, balancing fast inference and high accuracy.
OpenSubstance: A High-quality Measured Dataset of Multi-View and -Lighting Images and Shapes
Fan Pei (Zhejiang University), Hongzhi Wu (Zhejiang University)
SegmentationData SynthesisImageBenchmark
🎯 What it does: This paper constructs the OpenSubstance dataset, which contains 2.4 million HDR images of 187 real objects, covering 270 viewpoints and 1,637 lighting conditions, and provides precise camera/light parameters, foreground segmentation, and high-precision 3D geometry.
OpenVision: A Fully-Open, Cost-Effective Family of Advanced Vision Encoders for Multimodal Learning
Xianhang Li (University of California Santa Cruz), Cihang Xie (University of California Santa Cruz)
RecognitionData SynthesisRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A fully open and scalable visual encoder series named OpenVision has been developed for multimodal large models.
OphCLIP: Hierarchical Retrieval-Augmented Learning for Ophthalmic Surgical Video-Language Pretraining
Ming Hu (Shanghai AI Laboratory), Zongyuan Ge (Monash University)
RecognitionObject DetectionRetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper develops the OphCLIP visual-language pre-training framework and the OphVL large-scale hierarchical structure ophthalmic surgery video-text dataset to enhance phase recognition and multi-tool detection in ophthalmic surgery.
Optical Model-Driven Sharpness Mapping for Autofocus in Small Depth-of-Field and Severe Defocus Scenarios
Chen-Liang Fan (Foxconn), Yuesheng Zhu (Peking University)
Depth EstimationOptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: An automatic focusing method that combines optical model sharpness metrics and deep learning is proposed, utilizing sharpness-distance mapping to directly estimate the focus position from severely defocused images.
Optimal Transport for Brain-Image Alignment: Unveiling Redundancy and Synergy in Neural Information Processing
Yang Xiao (University of Tulsa), Bo Hui (University of Arizona)
OptimizationTransformerContrastive LearningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposed an optimal transport (OT) based loss function for the problem of brain image alignment, and applied it to the global matching of fMRI and image embeddings;
OracleFusion: Assisting the Decipherment of Oracle Bone Script with Structurally Constrained Semantic Typography
Caoshuo Li (Xiamen University), Rongrong Ji (Anyang Normal University)
GenerationTransformerLarge Language ModelDiffusion modelTextMultimodality
🎯 What it does: The OracleFusion two-stage semantic typography framework is proposed to assist in interpreting oracle bone script. It first uses a multimodal large language model to analyze the glyphs of oracle bone characters and locate key components, and then employs a structure-constrained vector fusion method to generate semantically rich and structurally complete vector fonts.
Orchid: Image Latent Diffusion for Joint Appearance and Geometry Generation
Akshay Krishnan (Google DeepMind), Abhijit Kundu (Google DeepMind)
GenerationData SynthesisDepth EstimationKnowledge DistillationDiffusion modelImageMultimodality
🎯 What it does: This paper presents Orchid, a unified latent diffusion model that can generate corresponding color, depth, and normal maps from text or a single RGB image, achieving consistent 3D appearance and geometric generation.
OrderChain: Towards General Instruct-Tuning for Stimulating the Ordinal Understanding Ability of MLLM
Jinhong Wang (Zhejiang University), Jian Wu (Transvascular Implantation Devices Research Institute and Liangzhu Laboratory)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: A new prompting paradigm called OrderChain is proposed, which enhances the understanding and prediction capabilities of multimodal large language models in ordinal regression tasks by utilizing task-aware prompts, range-optimized Chain-of-Thought, and category recursive partitioning.
ORION: A Holistic End-to-End Autonomous Driving Framework by Vision-Language Instructed Action Generation
Haoyu Fu (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
Autonomous DrivingTransformerLarge Language ModelVision Language ModelVision-Language-Action ModelGenerative Adversarial NetworkMultimodality
🎯 What it does: An end-to-end autonomous driving framework called ORION is proposed, which achieves the unification of driving decision-making and trajectory planning through action generation guided by visual language.
OURO: A Self-Bootstrapped Framework for Enhancing Multimodal Scene Understanding
Tianrun Xu (Tsinghua University), Feng Chen (Tsinghua University)
RecognitionObject DetectionGenerationTransformerVision Language ModelTextMultimodality
🎯 What it does: Proposes the OURO self-starting framework, which generates self-annotated data through recursive sub-region descriptions and VQA to train models, enhancing multimodal scene understanding;
Ouroboros: Single-step Diffusion Models for Cycle-consistent Forward and Inverse Rendering
Shanlin Sun (University of California), Chenyu You (Stony Brook University)
GenerationData SynthesisComputational EfficiencyDiffusion modelImageVideo
🎯 What it does: A framework named Ouroboros is proposed, which utilizes two single-step diffusion models to achieve inverse rendering and forward rendering, and jointly trains them through cycle consistency.
OuroMamba: A Data-Free Quantization Framework for Vision Mamba
Akshat Ramachandran (Georgia Institute of Technology), Tushar Krishna (Georgia Institute of Technology)
ClassificationObject DetectionSegmentationData SynthesisTransformerContrastive LearningImage
🎯 What it does: A two-stage data-free quantization framework called OuroMamba is proposed, which includes self-supervised generation of synthetic data and dynamic mixed-precision quantization.
Outdoor Monocular SLAM with Global Scale-Consistent 3D Gaussian Pointmaps
Chong Cheng (Hong Kong University of Science and Technology), Hao Wang (Hong Kong University of Science and Technology)
Autonomous DrivingOptimizationGaussian SplattingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: A monocular outdoor SLAM framework S3PO-GS is proposed, integrating 3D Gaussian Splatting with self-consistent point cloud tracking and patch dynamic mapping to achieve accurate localization and high-quality view synthesis.
Outlier-Aware Post-Training Quantization for Image Super-Resolution
Hailing Wang (Northeastern University), Yun Fu (Northeastern University)
RestorationSuper ResolutionImage
🎯 What it does: This paper proposes an offline post-training quantization method for image super-resolution that achieves low-bit quantization without the need for retraining or high-quality supervision.
OV-SCAN: Semantically Consistent Alignment for Novel Object Discovery in Open-Vocabulary 3D Object Detection
Adrian Chow (University of Waterloo), Krzysztof Czarnecki (University of Waterloo)
Object DetectionAutonomous DrivingTransformerPrompt EngineeringVision Language ModelPoint Cloud
🎯 What it does: The OV-SCAN framework is proposed, utilizing the SC-NOD module to generate high-quality 3D annotations and filtering low-quality 2D-3D alignment pairs through selective alignment, combined with H2SA two-stage hierarchical alignment to achieve open vocabulary 3D object detection.
OV3D-CG: Open-vocabulary 3D Instance Segmentation with Contextual Guidance
Mingquan Zhou (Institute of Computing Technology, Chinese Academy of Sciences), Xilin Chen (Institute of Computing Technology, Chinese Academy of Sciences)
Object DetectionSegmentationLarge Language ModelSimultaneous Localization and MappingMultimodalityPoint CloudChain-of-Thought
🎯 What it does: This paper proposes a context-guided open vocabulary 3D instance segmentation framework, OV3D-CG, which integrates a 3D pre-trained model, SAM 2D segmentation, and a multi-modal large language model (MLLM) for Chain-of-Thought reasoning, achieving precise segmentation and labeling of unknown category objects.
OVA-Fields: Weakly Supervised Open-Vocabulary Affordance Fields for Robot Operational Part Detection
Heng Su (Chongqing University), Chao Chen (Chongqing University)
Object DetectionRobotic IntelligenceTransformerGenerative Adversarial NetworkContrastive LearningImagePoint Cloud
🎯 What it does: The OVA-Fields framework is proposed, achieving open vocabulary 3D scene manipulation component perception and localization based on natural language instructions.
Overcoming Dual Drift for Continual Long-Tailed Visual Question Answering
Feifei Zhang (Tianjin University of Technology), Changsheng Xu (National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences)
RecognitionTransformerVision Language ModelMultimodality
🎯 What it does: To address the issues of continual learning and long-tail distribution in visual question answering, the CLT-VQA framework is proposed, which resolves intra-class prototype drift and inter-class feature drift through a dual balance of BCL and MFA.
OVG-HQ: Online Video Grounding with Hybrid-modal Queries
Runhao Zeng (Shenzhen MSU-BIT University), Xiping Hu (Shenzhen MSU-BIT University)
RetrievalKnowledge DistillationRepresentation LearningTransformerContrastive LearningImageVideoTextMultimodality
🎯 What it does: This paper proposes the Online Video Localization and Hybrid Query task (OVG-HQ) and implements a unified model that supports real-time temporal localization for multimodal queries such as text, images, and video segments.
p-AVAS: Can Physics-Integrated Audio-Visual Modeling Boost Neural Acoustic Synthesis?
Susan Liang (University of Rochester), Chenliang Xu (University of Rochester)
GenerationData SynthesisFlow-based ModelNeural Radiance FieldMultimodalityPhysics RelatedAudio
🎯 What it does: A two-stage physically integrated audio-visual synthesis framework π-AVAS is designed, which first generates rough audio through visual-driven 3D reconstruction and acoustic physical simulation, and then refines the simulated audio using a flow matching model to enhance realism.
p-MoD: Building Mixture-of-Depths MLLMs via Progressive Ratio Decay
Jun Zhang (Nanjing University), Limin Wang (Nanjing University)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: The p-MoD architecture is proposed, utilizing the Mixture-of-Depths mechanism to layer-wise filter visual tokens in the Transformer layer, significantly reducing the computational cost of multimodal large language models.
PacGDC: Label-Efficient Generalizable Depth Completion with Projection Ambiguity and Consistency
Haotian Wang (Xi'an Jiaotong University), Shijian Lu (Nanyang Technological University)
Data SynthesisDepth EstimationDomain AdaptationImagePoint Cloud
🎯 What it does: A high-efficiency label method for deep completion, PacGDC, has been developed, which significantly improves cross-domain generalization performance by generating pseudo-depth labels through projection ambiguity and consistency.
PAN-Crafter: Learning Modality-Consistent Alignment for PAN-Sharpening
Jeonghyeok Do (Korea Advanced Institute of Science and Technology), Munchurl Kim (Korea Advanced Institute of Science and Technology)
Image TranslationRestorationConvolutional Neural NetworkImage
🎯 What it does: A PAN-Crafter framework is proposed to fuse high-resolution panchromatic images with low-resolution multispectral images to generate high-resolution multispectral images.
PanoLlama: Generating Endless and Coherent Panoramas with Next-Token-Prediction LLMs
Teng Zhou (Zhejiang University), Yongchuan Tang (Zhejiang University)
GenerationData SynthesisTransformerLarge Language ModelText
🎯 What it does: Through autoregressive models and token redirection, PanoLlama is proposed to achieve continuous generation of untrained, infinitely long panoramic images.
PanoSplatt3R: Leveraging Perspective Pretraining for Generalized Unposed Wide-Baseline Panorama Reconstruction
Jiahui Ren (Northwestern Polytechnical University), Yuchao Dai (Northwestern Polytechnical University)
RestorationGenerationTransformerGaussian SplattingImage
🎯 What it does: A panoramic image dual-view wide baseline reconstruction and novel view synthesis framework called PanoSplatt3R has been developed without the need for pose information.
PanSt3R: Multi-view Consistent Panoptic Segmentation
Lojze Zust (Naver Labs Europe), Gabriela Csurka (Naver Labs Europe)
Object DetectionSegmentationTransformerImage
🎯 What it does: The PanSt3R framework is proposed, which can simultaneously achieve 3D reconstruction and multi-view consistent panoramic segmentation through a single forward inference using only an uncalibrated image set.
Parameter-Efficient Adaptation of Geospatial Foundation Models through Embedding Deflection
Romain Thoreau (National Centre for Space Studies), Dawa Derksen (National Centre for Space Studies)
ClassificationSegmentationTransformerImage
🎯 What it does: This paper proposes the DEFLECT method, which achieves parameter-efficient adaptation of RGB pre-trained geospatial foundation models (GFM) on multispectral satellite images.
Parametric Shadow Control for Portrait Generation in Text-to-Image Diffusion Models
Haoming Cai (University of Maryland), Christopher Metzler (University of Maryland)
GenerationData SynthesisDepth EstimationDiffusion modelImage
🎯 What it does: Developed Shadow Director, which allows for parametric and intuitive control of portrait shadows during the text-to-image diffusion model generation process, while maintaining consistency in identity and artistic style.
PARTE: Part-Guided Texturing for 3D Human Reconstruction from a Single Image
Hyeongjin Nam (Seoul National University), Kyoung Mu Lee (Seoul National University)
RestorationSegmentationGenerationTransformerDiffusion modelScore-based ModelImageMesh
🎯 What it does: This paper proposes a method for recovering high-quality 3D human geometry and texture from a single image, utilizing body part information to achieve precise texture alignment.
PartField: Learning 3D Feature Fields for Part Segmentation and Beyond
Minghua Liu (NVIDIA), Jun Gao (University of Toronto)
SegmentationTransformerContrastive LearningPoint CloudMesh
🎯 What it does: This paper presents PARTFIELD, a forward 3D model for learning continuous part feature fields, enabling tasks such as multi-scale part segmentation, co-segmentation, and correspondence.
Partial Forward Blocking: A Novel Data Pruning Paradigm for Lossless Training Acceleration
Dongyue Wu (Huazhong University of Science and Technology), Changxin Gao (Huazhong University of Science and Technology)
ClassificationSegmentationComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: A data pruning framework named Partial Forward Blocking (PFB) is proposed, which evaluates sample importance in shallow networks during training and prunes low-importance samples before deep networks to save computational costs in forward and backward propagation.
Partially Matching Submap Helps: Uncertainty Modeling and Propagation for Text to Point Cloud Localization
Mingtao Feng (Xidian University), Yaonan Wang (Hunan University)
Pose EstimationRetrievalAutonomous DrivingContrastive LearningTextMultimodalityPoint Cloud
🎯 What it does: The researchers proposed a more realistic scenario in the text-point cloud cross-modal localization task, allowing for partial matching between text and submaps, and improved localization accuracy through uncertainty modeling and propagation.
PASD: A Pixel-Adaptive Swarm Dynamics Approach for Unsupervised Low-Light Image Enhancement
Shuai Jin (Shanxi University), Xinyan Liang (Shanxi University)
RestorationReinforcement LearningImage
🎯 What it does: A pixel adaptive low-light image enhancement method PASD based on group dynamics is proposed, which achieves local detail and global brightness balance through dynamic neighborhood interactions between pixels.
PASG: A Closed-Loop Framework for Automated Geometric Primitive Extraction and Semantic Anchoring in Robotic Manipulation
Zhihao Zhu (Shanghai Jiao Tong University), Yao Mu (Shanghai Jiao Tong University)
Robotic IntelligenceTransformerSupervised Fine-TuningVision Language ModelMultimodalityPoint CloudBenchmark
🎯 What it does: A closed-loop framework called PASG is proposed, which automatically extracts geometric primitives through visual foundational models and binds them with functional semantics, enhancing the semantic level of robotic grasping and manipulation.
Passing the Driving Knowledge Test
Maolin Wei (Boston University), Eshed Ohn-Bar (Boston University)
Autonomous DrivingTransformerLarge Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the DriveQA benchmark to evaluate the knowledge question-answering ability of LLMs/MLLMs in traffic rules, signs, and right-hand driving.
PASTA: Part-Aware Sketch-to-3D Shape Generation with Text-Aligned Prior
Seunggwan Lee (Korea University), Sangpil Kim (Korea University)
GenerationData SynthesisGraph Neural NetworkTransformerVision Language ModelImagePoint Cloud
🎯 What it does: A framework called PASTA is proposed, which combines hand-drawn sketches and text priors for 3D shape generation and local editing.
PatchScaler: An Efficient Patch-Independent Diffusion Model for Image Super-Resolution
Yong Liu (Xi'an Jiaotong University), Fei Wang (Xi'an Jiaotong University)
RestorationSuper ResolutionTransformerDiffusion modelImage
🎯 What it does: This paper proposes PatchScaler, a single-image super-resolution method based on diffusion models, which can adaptively allocate sampling steps for different image patches and utilize texture prompts to enhance detail recovery.
PathDiff: Histopathology Image Synthesis with Unpaired Text and Mask Conditions
Mahesh Bhosale (University at Buffalo), Xuan Gong (Harvard Medical School)
SegmentationGenerationData SynthesisConvolutional Neural NetworkDiffusion modelImageText
🎯 What it does: A diffusion model called PathDiff has been developed to jointly generate pathological images from unpaired diagnostic texts and spatial masks.
PathFinder: A Multi-Modal Multi-Agent System for Medical Diagnostic Decision-Making Applied to Histopathology
Fatemeh Ghezloo (University of Washington), Linda Shapiro (University of Washington)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelAgentic AIImageMultimodalityBiomedical Data
🎯 What it does: PathFinder is proposed, a multimodal multi-agent system that simulates the navigation, description, and diagnostic processes of pathologists in whole slide images;
PBCAT: Patch-Based Composite Adversarial Training against Physically Realizable Attacks on Object Detection
Xiao Li (Tsinghua University), Xiaolin Hu (Tsinghua University)
Object DetectionAdversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: A Patch-Based Composite Adversarial Training (PBCAT) method is proposed to enhance the robustness of object detection against physically realizable attacks (adversarial patches and textures).
PBFG: A New Physically-Based Dataset and Removal of Lens Flares and Glares
Jie Zhu (Sungkyunkwan University), Sungkil Lee (Sungkyunkwan University)
RestorationData SynthesisTransformerImage
🎯 What it does: A PBFG physical benchmark glare and glare dataset is proposed, and based on this, the FGRNet network is trained to remove lens glare and glare in nighttime photography.
PCR-GS: COLMAP-Free 3D Gaussian Splatting via Pose Co-Regularizations
Yu Wei (Nanyang Technological University), Shijian Lu (University of Chinese Academy of Sciences)
Pose EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: A COLMAP-free 3D Gaussian scattering method PCR-GS is proposed, which optimizes camera poses and 3D Gaussians simultaneously through pose co-regularization.
PEFTDiff: Diffusion-Guided Transferability Estimation for Parameter-Efficient Fine-Tuning
Prafful Kumar Khoba (UQ-IITD Research Academy), Mahsa Baktashmotlagh
Diffusion modelImage
🎯 What it does: This paper proposes a diffusion graph-based metric for rapid and accurate ranking of various Parameter-Efficient Fine-Tuning (PEFT) methods under a shared backbone.
Penalizing Boundary Activation for Object Completeness in Diffusion Models
Haoyang Xu (Wuhan University), Yutian Lin (Wuhan University)
GenerationData SynthesisOptimizationTransformerDiffusion modelImage
🎯 What it does: This study addresses the issue of incomplete objects in images generated by diffusion models and proposes a lightweight method that is training-independent and only penalizes image boundary activations during the inference phase to enhance object completeness.
Perceive, Understand and Restore: Real-World Image Super-Resolution with Autoregressive Multimodal Generative Models
Hongyang Wei (Tsinghua University), Lei Zhang (Hong Kong Polytechnic University)
RestorationSuper ResolutionTransformerLarge Language ModelGenerative Adversarial NetworkImage
🎯 What it does: Developed the Real-ISR framework PURE based on the autoregressive multimodal generative model (Lumina-mGPT), which can perceive the degradation level of low-quality images, understand their semantic content, and achieve high-quality image generation and restoration within a single model.
Perceiving and Acting in First-Person: A Dataset and Benchmark for Egocentric Human-Object-Human Interactions
Liang Xu (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
Data SynthesisPose EstimationLarge Language ModelVision-Language-Action ModelVideoMultimodalityBenchmark
🎯 What it does: The InterVLA dataset, a large-scale human-object-human interaction dataset, is proposed, and four benchmark tasks are defined on it: human motion estimation, interaction synthesis, interaction prediction, and more.
Perception-as-Control: Fine-grained Controllable Image Animation with 3D-aware Motion Representation
Yingjie Chen (Alibaba Group), Liefeng Bo (Alibaba Group)
Object TrackingGenerationDiffusion modelVideo
🎯 What it does: This paper proposes a 3D-aware motion representation (simplifying the scene into key object spheres and a world envelope) and constructs a Perception-as-Control framework based on this representation. It utilizes perception results as motion control signals for spatial alignment, achieving fine-grained collaborative control of camera and object movements. Additionally, a three-stage training strategy is provided to balance camera and object movements.
Performing Defocus Deblurring by Modeling its Formation Process
Zhengbo Zhang (Singapore University of Technology and Design), De Wen Soh (Singapore University of Technology and Design)
RestorationTransformerImage
🎯 What it does: This paper proposes a single-image defocus deblurring (SIDD) method based on 2D Gaussian blur point (Illuminated Blob) representation, utilizing differentiable rasterization to convert defocused images into a set of adjustable 2D Gaussian blur points. A Transformer-based blob deblurrer is employed to adjust the scale, shape, and opacity of the blur points to achieve deblurring.
PerLDiff: Controllable Street View Synthesis Using Perspective-Layout Diffusion Model
Jinhua Zhang (University of Electronic Science and Technology of China), Shuhang Gu (University of Electronic Science and Technology of China)
Object DetectionSegmentationGenerationData SynthesisAutonomous DrivingDiffusion modelImage
🎯 What it does: Proposes PerLDiff, which achieves controllable street scene image generation based on 3D annotations through a perspective layout diffusion model.
PERSONA: Personalized Whole-Body 3D Avatar with Pose-Driven Deformations from a Single Image
Geonhee Sim (Korea University), Gyeongsik Moon (Korea University)
GenerationPose EstimationDiffusion modelGaussian SplattingImageVideo
🎯 What it does: This paper proposes a framework named PERSONA, which generates personalized full-body 3D avatars from a single image. It produces pose-rich videos through a diffusion model, combines SMPL-X with a mixed representation of 3D Gaussian Splatting, uses an MLP to predict the mean offset of each Gaussian for pose-driven deformation, and maintains identity consistency and rendering clarity through balanced sampling and geometric weighting optimization.
PersonaCraft: Personalized and Controllable Full-Body Multi-Human Scene Generation Using Occlusion-Aware 3D-Conditioned Diffusion
Gwanghyun Kim (Seoul National University), Se Young Chun (Seoul National University)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: PersonaCraft has been developed, a multi-person full-body personalization image synthesis framework based on SMPLx-Conditioned Diffusion, capable of generating realistic images with consistent poses and identity preservation in complex occlusion environments.
Personalized Federated Learning under Local Supervision
Qiqi Liu (Westlake University), Han Yu (Sony AI)
Federated LearningImage
🎯 What it does: This paper proposes a personalized federated learning framework FedSimSup based on a local supervisor's parallel structure to improve model performance under non-IID conditions.
PersonalVideo: High ID-Fidelity Video Customization without Dynamic and Semantic Degradation
Hengjia Li (Zhejiang University), Deng Cai (Zhejiang University)
GenerationData SynthesisReinforcement LearningDiffusion modelVideo
🎯 What it does: This paper proposes a personalized video generation framework called PersonalVideo, which directly applies reward supervision on the generated videos, avoiding the parameter tuning inference gap caused by traditional reconstruction-based training, while maintaining high identity fidelity without sacrificing motion dynamics and semantic consistency.
Perspective-aware 3D Gaussian Inpainting with Multi-view Consistency
Yuxin Cheng (University of Hong Kong), Ngai Wong (University of Hong Kong)
RestorationDepth EstimationDiffusion modelGaussian SplattingImage
🎯 What it does: This paper proposes a perspective-aware 3D Gaussian inpainting framework called PAInpainter, which achieves high-quality, cross-view consistent inpainting in multi-view scenes.
Perspective-Aware Reasoning in Vision-Language Models via Mental Imagery Simulation
Phillip Y. Lee (KAIST), Minhyuk Sung (KAIST)
Object DetectionSegmentationDepth EstimationTransformerVision Language ModelImageTextMultimodality
🎯 What it does: By constructing scene abstractions and implementing perspective transformations, the APC framework is proposed to assist visual language models (VLM) in spatial reasoning from any reference perspective.
Perspective-Aware Teaching: Adapting Knowledge for Heterogeneous Distillation
Jhe-Hao Lin (National Yang Ming Chiao Tung University), Wen-Huang Cheng (National Taiwan University)
Object DetectionKnowledge DistillationConvolutional Neural NetworkTransformerPrompt EngineeringImage
🎯 What it does: This paper proposes a Perspective-Aware Teacher (PAT) framework for knowledge distillation across different architectures (CNN, ViT, MLP).
Perspective-Invariant 3D Object Detection
Ao Liang (National University of Singapore), Wei Tsang Ooi (National University of Singapore)
Object DetectionAutonomous DrivingPoint Cloud
🎯 What it does: This paper proposes Pi3DET, the first LiDAR 3D detection dataset covering multiple platforms such as vehicles, drones, and quadruped robots, and presents a cross-platform adaptive framework Pi3DET-Net.
PersPose: 3D Human Pose Estimation with Perspective Encoding and Perspective Rotation
Xiaoyang Hao (Southern University of Science and Technology), Han Li (Southern University of Science and Technology)
Pose EstimationConvolutional Neural NetworkImage
🎯 What it does: This paper presents PersPose, a 3D human pose estimation framework that integrates perspective encoding and perspective rotation.
Ph-GAN: Physics-Inspired GAN for Generating SAR Images Under Limited Data
Xidan Zhang (Northwestern Polytechnical University), Zhongling Huang (Northwestern Polytechnical University)
GenerationData SynthesisGenerative Adversarial NetworkImagePhysics Related
🎯 What it does: This paper proposes a physics-based Generative Adversarial Network (Φ-GAN) for generating synthetic aperture radar (SAR) images under data-scarce conditions, capable of producing images with realistic scattering characteristics from unseen viewpoints.
Phantom: Subject-Consistent Video Generation via Cross-Modal Alignment
Lijie Liu (ByteDance), Xinglong Wu (ByteDance)
GenerationData SynthesisTransformerVision Language ModelRectified FlowImageVideoText
🎯 What it does: The Phantom framework is proposed, which generates videos consistent with reference images and compliant with text descriptions through cross-modal alignment (text, image, video triplets), supporting both single and multiple subject scenarios.
PHATNet: A Physics-guided Haze Transfer Network for Domain-adaptive Real-world Image Dehazing
Fu-Jen Tsai (National Tsing Hua University), Chia-Wen Lin (National Tsing Hua University)
RestorationDomain AdaptationConvolutional Neural NetworkImagePhysics Related
🎯 What it does: Utilizing a physics-guided haze transfer network (PHATNet) for haze pattern migration between the source domain and the target domain, generating domain-specific fine-tuning datasets to achieve adaptive enhancement for real-time image dehazing.
PHD: Personalized 3D Human Body Fitting with Point Diffusion
Hsuan-I Ho (ETH Zurich), Linguang Zhang (Meta)
Pose EstimationOptimizationTransformerRectified FlowVideoPoint Cloud
🎯 What it does: This paper proposes a personalized 3D human pose and shape recovery framework called PHD, which first calibrates the user's body shape using a single-frame reference pose, and then provides a 3D prior for pose fitting through a shape-based Point Diffusion Transformer (PointDiT), forming a point distillation sampling-guided optimization.
Photolithography Overlay Map Generation with Implicit Knowledge Distillation Diffusion Transformer
Yuan-Fu Yang (National Yang Ming Chiao Tung University), Hsiu-Hui Hsiao (National Taiwan University of Science and Technology)
GenerationKnowledge DistillationTransformerDiffusion modelAuto EncoderContrastive LearningImageMultimodality
🎯 What it does: Proposes the IKDDiT framework, which utilizes implicit knowledge distillation, diffusion transformers, and unified contrastive learning to generate photolithography overlay images.
Physical Degradation Model-Guided Interferometric Hyperspectral Reconstruction with Unfolding Transformer
Yuansheng Li (Beijing Institute of Technology), Ying Fu (Beijing Institute of Technology)
RestorationData SynthesisTransformerImage
🎯 What it does: This paper proposes an interference-based hyperspectral imaging reconstruction method based on a physical attenuation model—IHRUT—and generates realistic synthetic data using this model.
Physics Context Builders: A Modular Framework for Physical Reasoning in Vision-Language Models
Vahid Balazadeh (University of Toronto), Rahul G. Krishnan (University of Toronto)
TransformerLarge Language ModelVision Language ModelImageTextPhysics Related
🎯 What it does: The Physics Context Builders (PCBs) framework is proposed, which enhances the physical reasoning capabilities of large VLMs by training small VLMs to generate physical scene descriptions in a simulated environment.
PhysRig: Differentiable Physics-Based Skinning and Rigging Framework for Realistic Articulated Object Modeling
Hao Zhang (University of Illinois Urbana Champaign), Narendra Ahuja (University of Illinois Urbana Champaign)
GenerationData SynthesisPose EstimationMeshPhysics Related
🎯 What it does: This paper presents PhysRig, a differentiable physics-driven skinning and binding framework for realistic simulation of pose variations in deformable organic and flexible bodies.
PhysSplat: Efficient Physics Simulation for 3D Scenes via MLLM-Guided Gaussian Splatting
Haoyu Zhao (Wuhan University), Hua Zou (Wuhan University)
GenerationComputational EfficiencyLarge Language ModelGaussian SplattingMultimodalityPoint CloudPhysics Related
🎯 What it does: By combining 3D reconstruction, object-level 3D open vocabulary segmentation, zero-shot learning with a multimodal large language model (MLLM), predicting physical properties, distribution estimation networks, and physically-based geometry adaptive sampling, we achieve real-time generation of 3D dynamic scenes with realistic physical interactions on a single GPU.