ICCV 2025 Papers — Page 20
IEEE/CVF International Conference on Computer Vision · 2701 papers
RA-BUSSeg: Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment
Wanting Zhang (Shenzhen University), Jing Qin (Hong Kong Polytechnic University)
SegmentationConvolutional Neural NetworkTransformerImageBiomedical DataUltrasound
🎯 What it does: A semi-supervised thymus ultrasound image segmentation model RA-BUSSeg is proposed, combining two main modules: Adjacent Pixel Relationship Propagation (ARP) and Cross-layer Relationship Alignment (CRA) to enhance segmentation quality by utilizing pixel-level relationships.
RadarSplat: Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes
Pou-Chun Kung (University of Michigan), Katherine A. Skinner (University of Michigan)
Data SynthesisAutonomous DrivingGaussian SplattingPoint Cloud
🎯 What it does: Developed RadarSplat, which uses Gaussian Splatting to render radar images and perform 3D reconstruction and occupancy estimation.
RadGPT: Constructing 3D Image-Text Tumor Datasets
Pedro R.A.S. Bassi, Zongwei Zhou (Johns Hopkins University)
SegmentationGenerationConvolutional Neural NetworkLarge Language ModelImageTextMultimodalityBiomedical DataComputed TomographyBenchmark
🎯 What it does: This study presents AbdomenAtlas 3.0, a publicly available large-scale abdominal CT dataset containing 9,262 3D CT scans, tumor annotations for each voxel, and corresponding radiology reports. Additionally, the RadGPT framework was developed to convert tumor segmentation results into structured and narrative reports, enabling automated report generation through LLM. Benchmark evaluations of six CT report generation models on this dataset demonstrate that the segmentation-based RadGPT significantly outperforms end-to-end models in tumor detection sensitivity and specificity.
Radiant Foam: Real-Time Differentiable Ray Tracing
Shrisudhan Govindarajan (Simon Fraser University), Andrea Tagliasacchi (University of Toronto)
OptimizationComputational EfficiencyNeural Radiance FieldGaussian SplattingPoint Cloud
🎯 What it does: A differentiable real-time ray tracing scene representation called Radiant Foam is proposed, utilizing the foam structure of Voronoi volume grids to achieve a learnable light field.
RAGD: Regional-Aware Diffusion Model for Text-to-Image Generation
Zhennan Chen (Nanjing University), Ying Tai (Nanjing University)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: This paper proposes the RAGD (Regional-Aware Diffusion Model), which splits the original prompt into regional prompts. It first constructs each region using Regional Hard Binding in the early steps, and then integrates regional information using Regional Soft Refinement in the later steps, achieving precise control over location, attributes, and relationships, while supporting regional repainting of generated images without introducing additional patching models.
RAGDiffusion: Faithful Cloth Generation via External Knowledge Assimilation
Yuhan Li (Shanghai Jiao Tong University), Bingbing Ni (Shanghai Jiao Tong University)
GenerationData SynthesisDiffusion modelContrastive LearningImageRetrieval-Augmented Generation
🎯 What it does: A RAGDiffusion framework based on Retrieval-Augmented Generation (RAG) is proposed for generating front-view flat clothing images from real scenes.
RAGNet: Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping
Dongming Wu (Chinese University of Hong Kong), Jianbing Shen (University of Macau)
SegmentationRobotic IntelligenceTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageBenchmark
🎯 What it does: A large-scale semantic reasoning-based grasping feasibility segmentation benchmark RAGNet is constructed, and AffordanceNet is proposed to achieve feasibility segmentation and 3D grasping under language conditions through VLM and a pose generator.
RainbowPrompt: Diversity-Enhanced Prompt-Evolving for Continual Learning
Kiseong Hong (Chung Ang University), Eunwoo Kim
ClassificationRecognitionTransformerPrompt EngineeringImageVideo
🎯 What it does: A prompt-evolving mechanism (RainbowPrompt) is proposed, which dynamically aggregates and enhances the diversity of task-specific prompts through self-attention transformation and task-guided alignment to achieve continual learning without replay.
RALoc: Enhancing Outdoor LiDAR Localization via Rotation Awareness
Yuyang Yang (Xiamen University), Cheng Wang (Xiamen University)
Pose EstimationAutonomous DrivingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: A rotation-aware framework for LiDAR localization, RALoc, is proposed, which combines point cloud normalization and scene coordinate regression.
Randomized Autoregressive Visual Generation
Qihang Yu (ByteDance), Liang-Chieh Chen (ByteDance)
GenerationData SynthesisTransformerImage
🎯 What it does: This paper proposes a Randomized Autoregressive (RAR) model for image generation.
RANKCLIP: Ranking-Consistent Language-Image Pretraining
Yiming Zhang (University of Science and Technology of China), Yining Sun (Chinese Academy of Sciences)
ClassificationRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A language-image pre-training method based on list-level ranking consistency, RANKCLIP, is proposed, enhancing the alignment process of CLIP and capable of capturing multi-to-multi relationships across modalities and within a single modality.
RapVerse: Coherent Vocals and Whole-Body Motion Generation from Text
Jiaben Chen (University of Massachusetts Amherst), Chuang Gan (MIT IBM Watson AI Lab)
GenerationData SynthesisTransformerAuto EncoderTextMultimodalityAudio
🎯 What it does: This paper proposes a unified text-conditioned multimodal generation framework that can simultaneously generate synthesized singing voice and full-body 3D motion from rap lyrics.
RARE: Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning
Chengyu Zheng (Nanjing University of Aeronautics and Astronautics), Mingqiang Wei (Nanjing University of Aeronautics and Astronautics)
OptimizationDiffusion modelPoint Cloud
🎯 What it does: A method for zero-shot refinement of point cloud registration is proposed, utilizing a pre-trained diffusion model to extract diffusion features from depth maps, which are fused with existing geometric features to improve registration accuracy.
RareCLIP: Rarity-aware Online Zero-shot Industrial Anomaly Detection
Jianfang He (Institute of Automation, Chinese Academy of Sciences), Qiong Xie (Institute of Automation, Chinese Academy of Sciences)
Anomaly DetectionTransformerVision Language ModelImage
🎯 What it does: The RareCLIP framework is proposed for online zero-shot industrial anomaly detection.
RayGaussX: Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis
Hugo Blanc (Mines Paris), Alexis Paljic (Mines Paris)
GenerationData SynthesisComputational EfficiencyNeural Radiance FieldGaussian SplattingImage
🎯 What it does: Based on RayGauss, RayGaussX is proposed, which significantly accelerates the training and inference of Gaussian voxel ray marching through empty space skipping, adaptive sampling, ray and Gaussian reordering, scale regularization, and improved densification criteria.
RayletDF: Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians
Shenxing Wei (Hong Kong Polytechnic University), Bo Yang (Hong Kong Polytechnic University)
SegmentationGenerationDepth EstimationConvolutional Neural NetworkSupervised Fine-TuningPoint Cloud
🎯 What it does: A general 3D surface reconstruction method called RayletDF is proposed, which directly predicts surface points using raylet distance fields, supporting point cloud or 3D Gaussian data input, and can obtain a complete scene surface with a single forward inference.
RayPose: Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation
Junwen Huang, Benjamin Busam (Technical University of Munich)
Pose EstimationTransformerDiffusion modelImage
🎯 What it does: This paper proposes a template view Ray Bundling Diffusion method based on a multi-view diffusion model to achieve 6D pose estimation of unseen objects from a single RGB image.
RayZer: A Self-supervised Large View Synthesis Model
Hanwen Jiang (University of Texas at Austin), Georgios Pavlakos (University of Texas at Austin)
Data SynthesisPose EstimationTransformerVideo
🎯 What it does: This paper presents RayZer, a multi-view 3D vision model that can be trained completely without 3D supervision (camera pose or scene geometry), capable of performing camera estimation, scene reconstruction, and novel view synthesis on images without pose annotations.
RCTDistill: Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion
Geonho Bang (Seoul National University), Jun Won Choi (Seoul National University)
Object DetectionAutonomous DrivingKnowledge DistillationMultimodalityPoint Cloud
🎯 What it does: A cross-modal knowledge distillation framework (RCTDistill) is proposed to transfer knowledge from a LiDAR teacher to a radar-camera student model, enhancing 3D object detection accuracy through temporal fusion.
ReAL-AD: Towards Human-Like Reasoning in End-to-End Autonomous Driving
Yuhang Lu (ShanghaiTech University), Xinge Zhu (Chinese University of Hong Kong)
Autonomous DrivingTransformerVision Language ModelMultimodality
🎯 What it does: This study proposes an end-to-end autonomous driving framework named ReAL-AD, which utilizes a vision-language model for semantic reasoning at three levels of human cognition (strategy, decision-making, operation) in driving scenarios, and embeds it into a planning network.
Real3D: Towards Scaling Large Reconstruction Models with Real Images
Hanwen Jiang (University of Texas at Austin), Georgios Pavlakos (University of Texas at Austin)
RestorationSegmentationDepth EstimationContrastive LearningImage
🎯 What it does: A large reconstruction model is trained on single-view real images through a self-supervised framework, addressing the reliance on multi-view supervision.
RealCam-I2V: Real-World Image-to-Video Generation with Interactive Complex Camera Control
Teng Li (Zhejiang University), Xi Li (Huawei)
GenerationData SynthesisDepth EstimationDiffusion modelImageVideoPoint Cloud
🎯 What it does: Designed RealCam-I2V, a framework for generating videos from images, which utilizes camera trajectories to guide video generation, addressing issues of camera scale inconsistency and usability.
RealGeneral: Unifying Visual Generation via Temporal In-Context Learning with Video Models
Yijing Lin (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderVideoMultimodality
🎯 What it does: RealGeneral unifies multimodal image generation tasks as the next frame prediction task in video models, utilizing a pre-trained video diffusion model to generate various image conditional scenes in one go;
Reangle-A-Video: 4D Video Generation as Video-to-Video Translation
Hyeonho Jeong (KAIST), Jong Chul Ye (KAIST)
GenerationData SynthesisSupervised Fine-TuningDiffusion modelVideo
🎯 What it does: Using single-view videos, a two-stage framework is proposed (first, self-supervised fine-tuning of the video diffusion model to learn viewpoint-invariant motion through warped videos, and then generating multi-view consistent starting frames via warp-and-inpaint), achieving the generation of synchronized multi-view 4D videos from input videos;
ReasonVQA: A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering
Duong T. Tran (Bosch), Danh Le Phuoc (Technical University of Berlin)
RecognitionRetrievalTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: A large-scale multi-hop knowledge-driven visual question answering dataset, ReasonVQA, has been constructed, along with a low-cost scalable automatic generation framework.
ReassembleNet: Learnable Keypoints and Diffusion for 2D Fresco Reconstruction
Adeela Islam (Fondazione Istituto Italiano di Tecnologia), Alessio Del Bue (Fondazione Istituto Italiano di Tecnologia)
RestorationPose EstimationGraph Neural NetworkDiffusion modelImage
🎯 What it does: ReassembleNet proposes an end-to-end deep learning framework that learns keypoint selection, fuses multimodal features, and utilizes diffusion models for the reassembly of 2D fresco fragments.
ReCamMaster: Camera-Controlled Generative Rendering from A Single Video
Jianhong Bai (Kuaishou Tech), Di Zhang (Kuaishou Tech)
GenerationData SynthesisDiffusion modelVideo
🎯 What it does: ReCamMaster is a camera control video re-rendering framework based on a pre-trained text-to-video diffusion model, capable of generating new videos that conform to specified camera trajectories from a single input video.
Recognizing Actions from Robotic View for Natural Human-Robot Interaction
Ziyi Wang (Peking University), Mengyuan Liu (Peking University)
RecognitionRobotic IntelligenceTransformerVideoMultimodalityPoint Cloud
🎯 What it does: This work first constructs a large-scale active viewpoint human-robot interaction action dataset called ACTIVE, and proposes a new framework for action recognition from the robot's perspective, named ACTIVE-PC;
ReconDreamer++: Harmonizing Generative and Reconstructive Models for Driving Scene Representation
Guosheng Zhao (Institute of Automation Chinese Academy of Sciences), Xingang Wang (Institute of Automation Chinese Academy of Sciences)
GenerationAutonomous DrivingDiffusion modelNeural Radiance FieldGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes ReconDreamer++, a driving scene representation framework that integrates generative models with reconstruction models, capable of achieving high-quality rendering on both original and novel trajectories. It enhances consistency and detail by independently modeling the ground surface, introducing a spatial deformation network, and employing deep supervision.
ReCoT: Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models
Mengxue Qu (Beijing Jiaotong University), Yao Zhao (Beijing Jiaotong University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: This paper proposes a two-stage training framework called ReCoT, which first cultivates the model's self-reflection and error correction abilities in the Reflective Fine-Tuning stage, and then aligns the model's reflections with the final answers through Consistency Direct Preference Optimization, significantly reducing the confirmation bias of large visual-language models.
Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior
Renzhi He (University of California), Yi Xue (University of California)
RestorationNeural Radiance FieldImagePhysics Related
🎯 What it does: In the context of sparse perspectives, a Neural Volumetric Prior (NVP) mixed neural field model is proposed to recover the three-dimensional refractive index distribution of cells through fluorescence diffraction images.
Recovering Parametric Scenes from Very Few Time-of-Flight Pixels
Carter Sifferman (University of Wisconsin-Madison), Yin Li (University of Wisconsin-Madison)
Pose EstimationDepth EstimationOptimizationTransformerPoint Cloud
🎯 What it does: Recovering the geometry of simple parametric 3D scenes (such as the 6D pose of known objects) using only 15 low-cost single-pixel ToF sensors.
Rectifying Magnitude Neglect in Linear Attention
Qihang Fan (Institute of Automation, Chinese Academy of Sciences), Ran He (Institute of Automation, Chinese Academy of Sciences)
ClassificationObject DetectionSegmentationTransformerDiffusion modelImageAudio
🎯 What it does: A new linear attention mechanism called Magnitude-Aware Linear Attention (MALA) is proposed, and based on it, the MAViT visual transformer is constructed.
Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization
Xu Zheng (Hong Kong University of Science and Technology), Xuming Hu (Hong Kong University of Science and Technology)
SegmentationSupervised Fine-TuningImageMultimodality
🎯 What it does: A multi-scale method based on functional entropy regularization is proposed, which directly incorporates regularization at both the feature and prediction layers, aiming to suppress unimodal bias in multimodal semantic segmentation without the need for additional fusion modules.
REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents
Rui Tian (Fudan University), Yu-Gang Jiang (Fudan University)
GenerationCompressionComputational EfficiencyTransformerDiffusion modelAuto EncoderVideoText
🎯 What it does: Proposes Reducio-VAE and Reducio-DiT for efficient generation of high-resolution videos, compressing the video latent space to 1/64 of the original, significantly improving training and inference speed.
RefEdit: A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions
Bimsara Pathiraja (Arizona State University), Chitta Baral (Arizona State University)
GenerationData SynthesisLarge Language ModelDiffusion modelImageBenchmark
🎯 What it does: This paper presents RefEdit, an instruction-based image editing model specifically designed for precise editing of referring expressions.
Refer to Any Segmentation Mask Group With Vision-Language Prompts
Shengcao Cao (University of Illinois Urbana-Champaign), Yu-Xiong Wang (University of Illinois Urbana-Champaign)
Object DetectionSegmentationLarge Language ModelPrompt EngineeringImageMultimodality
🎯 What it does: The task of 'omnimodal referring expression segmentation (ORES)' is proposed, and the 'Refer to Any Segmentation Mask Group (RAS)' framework is developed, which can accept both text and visual (mask) multimodal prompts and output a set of masks that meet the prompts.
ReferDINO: Referring Video Object Segmentation with Visual Grounding Foundations
Tianming Liang (Sun Yat-sen University), Jian-Fang Hu (Sun Yat-sen University)
Object DetectionSegmentationTransformerSupervised Fine-TuningVideoText
🎯 What it does: In the video object segmentation task, a ReferDINO model based on GroundingDINO is proposed, achieving end-to-end text description-driven video object segmentation.
Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion
Byeonghun Lee (Korea University), Kyong Hwan Jin (Korea University)
RestorationSuper ResolutionRetrievalDiffusion modelImageRetrieval-Augmented Generation
🎯 What it does: A retrieval-augmented generation image super-resolution framework, iRAG, is proposed, which can automatically retrieve high-resolution reference images that match low-resolution inputs and perform diffusion recovery in the latent space.
ReferEverything: Towards Segmenting Everything We Can Speak of in Videos
Anurag Bagchi (Carnegie Mellon University), Martial Hebert (Carnegie Mellon University)
Object DetectionSegmentationDiffusion modelAuto EncoderVideoText
🎯 What it does: A framework named REM has been developed, which utilizes the visual-language mapping learned from a pre-trained video diffusion model to perform segmentation directly on videos based on natural language descriptions, supporting dynamic concepts of objects and non-objects.
Referring Expression Comprehension for Small Objects
Kanoko Goto (Institute of Science), Nakamasa Inoue (Institute of Science)
Object DetectionAutonomous DrivingTransformerPrompt EngineeringImage
🎯 What it does: For the task of referring expression comprehension (REC) for extremely small objects, the authors constructed a new SOREC dataset and proposed the Progressive‑Iterative Zooming Adapter (PIZA) module, enabling the model to progressively zoom in and locate the target in an autoregressive manner.
Referring to Any Person
Qing Jiang (South China University of Technology), Lei Zhang (International Digital Economy Academy)
RecognitionObject DetectionRetrievalTransformerLarge Language ModelImageMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposed the human multi-instance referencing task and its corresponding dataset HumanRef, along with the retrieval-based large language model RexSeek.
Reflect-DiT: Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection
Shufan Li (University of California Los Angeles), Aditya Grover (University of California Los Angeles)
Object DetectionGenerationData SynthesisTransformerVision Language ModelDiffusion modelImageText
🎯 What it does: Introducing a reflection mechanism during inference in text-to-image diffusion models allows the model to gradually improve generation results by utilizing previous images and natural language feedback during the generation process.
ReFlex: Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation
Jimyeong Kim (Seoul National University), Wonjong Rhee (Seoul National University)
Image TranslationGenerationDiffusion modelRectified FlowImage
🎯 What it does: A training-free, mask-free ReFlex method is proposed to achieve real image-text guided editing of the Rectified Flow (FLUX) model.
REGEN: Learning Compact Video Embedding with (Re-)Generative Decoder
Yitian Zhang (Adobe), Yun Fu (Northeastern University)
GenerationCompressionTransformerDiffusion modelVideoText
🎯 What it does: We propose REGEN, a video embedding design structured as an encoder-generator framework, utilizing a diffusion Transformer as the decoder to achieve up to 32× temporal compression while maintaining high-quality reconstruction and text-to-video generation.
RegGS: Unposed Sparse Views Gaussian Splatting with 3DGS Registration
Chong Cheng (Hong Kong University of Science and Technology), Hao Wang (Hong Kong University of Science and Technology)
GenerationOptimizationGaussian SplattingVideo
🎯 What it does: We propose RegGS, a registration framework based on 3D Gaussian splatting for reconstructing scenes from uncalibrated sparse views.
Region-aware Anchoring Mechanism for Efficient Referring Visual Grounding
Shuyi Ouyang (Zhejiang University), Lanfen Lin (Zhejiang University)
RecognitionObject DetectionComputational EfficiencyTransformerVision Language ModelImageMultimodality
🎯 What it does: Proposes a Region-aware Anchoring Mechanism (RaAM) that achieves efficient Referring Visual Grounding by learning region anchors that alternate and interact with visual and linguistic features.
Region-based Cluster Discrimination for Visual Representation Learning
Yin Xie (DeepGlint), Jiankang Deng (Imperial College London)
RecognitionObject DetectionRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: This paper proposes a Region-level Clustering Discrimination (RICE) framework, constructing a billion-level candidate region dataset, and incorporates a Region Transformer layer and a unified region clustering loss into the visual encoder to enhance the visual model's ability to perceive objects and OCR regions.
Region-Level Data Attribution for Text-to-Image Generative Models
Trong Bang Nguyen (Hanoi University of Science and Technology), Minh Hoai (Australian Institute of Machine Learning)
Object DetectionGenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: This paper proposes a region-level data attribution framework for text-to-image diffusion models, capable of locating the most influential image regions in training samples and providing attribution scores.
Registration beyond Points: General Affine Subspace Alignment via Geodesic Distance on Grassmann Manifold
Jaeho Shin (Seoul National University), Ayoung Kim (Seoul National University)
Pose EstimationOptimizationSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: This paper proposes an optimizable cost function for affine subspace alignment based on Grassmann distance, and implements a globally optimal line registration algorithm using a Branch-and-Bound (BnB) solver, significantly improving registration accuracy under noise and outlier conditions.
Reinforcement Learning-Guided Data Selection via Redundancy Assessment
Suorong Yang (Nanjing University), Jian Zhao (Nanjing University)
Computational EfficiencyData-Centric LearningConvolutional Neural NetworkTransformerReinforcement LearningImage
🎯 What it does: A dynamic data selection framework called RL-Selector based on reinforcement learning is proposed, which improves training efficiency and model generalization ability by evaluating sample redundancy and adaptively pruning unnecessary data during the training process.
Relative Illumination Fields: Learning Medium and Light Independent Underwater Scenes
Mengkun She (Kiel University), Kevin Köser (Kiel University)
RestorationData SynthesisNeural Radiance FieldImage
🎯 What it does: This study investigates a NeRF that models the light field in the local camera view and the scattering medium in an underwater scene under unknown, moving lighting, achieving clear scene reconstruction and viewpoint synthesis independent of lighting and medium.
ReME: A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation
Xiwei Xuan (University of California), Kwan-Liu Ma (University of California)
SegmentationData-Centric LearningContrastive LearningImage
🎯 What it does: A data-driven, training-free open vocabulary semantic segmentation framework called ReME is proposed, which utilizes high-quality segment texts to achieve retrieval-based segmentation for the reference set.
Reminiscence Attack on Residuals: Exploiting Approximate Machine Unlearning for Privacy
Yaxin Xiao (Hong Kong Polytechnic University), Yijie Jiao (Hong Kong Polytechnic University)
GenerationSafty and PrivacyConvolutional Neural NetworkReinforcement LearningDiffusion modelImage
🎯 What it does: This paper studies the implicit residuals generated during the Approximate Machine Unlearning (AMU) process and proves that these residuals can leak the privacy of forgotten data. It subsequently proposes the Reminiscence Attack (ReA) to exploit these residuals for membership inference attacks and designs the Orthogonal Unlearning & Replay (OUR) two-phase framework to eliminate the residuals while maintaining model performance.
Removing Cost Volumes from Optical Flow Estimators
Simon Kiefhaber (Technical University of Darmstadt), Simone Schaub-Meyer (Technical University of Darmstadt)
Computational EfficiencyConvolutional Neural NetworkOptical FlowImageVideo
🎯 What it does: During the training phase, the cost volume in the optical flow estimator is gradually removed, ultimately resulting in an efficient optical flow model (ReCoVEr) that does not rely on the cost volume.
Removing Out-of-Focus Reflective Flares via Color Alignment
Fengbo Lan (Hong Kong Polytechnic University), Chang Wen Chen (Hong Kong Polytechnic University)
Image HarmonizationRestorationDiffusion modelImage
🎯 What it does: This paper addresses the problem of removing defocused halo artifacts by transforming the halo removal task into color distribution alignment, using diffusion models and differentiable histogram loss for training, and directly performing detailed color correction on halo-affected images instead of traditional filling methods.
ReMP-AD: Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection
Hongchi Ma (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
Anomaly DetectionTransformerVision Language ModelImageMultimodality
🎯 What it does: This paper proposes a retrieval-enhanced multimodal prompt fusion framework, ReMP-AD, for industrial visual anomaly detection with few samples, which includes two main modules: Intra-class Token Retrieval (ICTR) and Visual-Language Prior Fusion (VLPF).
Rep-MTL: Unleashing the Power of Representation-level Task Saliency for Multi-Task Learning
Zedong Wang (Hong Kong University of Science and Technology), Dan Xu (Hong Kong University of Science and Technology)
SegmentationOptimizationContrastive LearningImage
🎯 What it does: This paper proposes a multi-task learning optimization method based on task saliency at the representation layer, called Rep-MTL, which reduces negative transfer and enhances task complementarity in the shared representation space through entropy regularization and cross-task alignment.
REPA-E: Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers
Xingjian Leng (Australian National University), Liang Zheng (Australian National University)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImage
🎯 What it does: This paper proposes the REPA-E end-to-end training strategy, which synchronously updates the Variational Autoencoder (VAE) and the Latent Diffusion Model (LDM), avoiding the latent space degradation caused by traditional two-stage training.
REPARO: Compositional 3D Assets Generation with Differentiable 3D Layout Alignment
Haonan Han (Tsinghua University), Wanhua Li (Harvard University)
GenerationOptimizationDiffusion modelImage
🎯 What it does: The REPARO two-stage process is proposed, which extracts multiple objects from a single image and generates 3D assets separately, then optimizes the object layout using differentiable rendering to obtain a complete multi-object 3D scene.
RePoseD: Efficient Relative Pose Estimation With Known Depth Information
Yaqing Ding (Comenius University in Bratislava), Zuzana Kukelova (Czech Technical University in Prague)
Pose EstimationDepth EstimationComputational EfficiencySimultaneous Localization and MappingImage
🎯 What it does: The research utilizes monocular depth estimation (MDE) results to propose a new relative pose solver for cameras that can simultaneously estimate scale/offset parameters and relative pose.
Representation Shift: Unifying Token Compression with FlashAttention
Joonmyung Choi (Korea University), Hyunwoo J. Kim (KAIST)
RetrievalCompressionConvolutional Neural NetworkTransformerImageVideo
🎯 What it does: This paper proposes a training-free, model-agnostic 'Representation Shift' metric to measure the extent of representation change for each token after MLP processing in Transformer layers. Based on this metric, it compresses tokens in video, image, and CNN/SSM models, compatible with FlashAttention for acceleration.
Representing 3D Shapes with 64 Latent Vectors for 3D Diffusion Models
In Cho (Yonsei University), Seon Joo Kim (Yonsei University)
GenerationCompressionComputational EfficiencyTransformerDiffusion modelAuto EncoderPoint CloudMesh
🎯 What it does: Proposes the COD-VAE two-stage autoencoder framework, which compresses 3D shapes into only 64 1D latent vectors, significantly improving generation efficiency and quality.
Repurposing 2D Diffusion Models with Gaussian Atlas for 3D Generation
Tiange Xiang (Stanford University), Li Fei-Fei (Stanford University)
GenerationData SynthesisDiffusion modelPoint CloudMesh
🎯 What it does: Using a pre-trained two-dimensional diffusion model, the generation of three-dimensional Gaussian objects from text is achieved by mapping a three-dimensional Gaussian distribution to a two-dimensional Gaussian Atlas.
RESCUE: Crowd Evacuation Simulation via Controlling SDM-United Characters
Xiaolin Liu (Tianjin University), Kun Li (Tianjin University)
Robotic IntelligenceDiffusion model
🎯 What it does: A real-time 3D crowd evacuation simulation framework (RESCUE) has been constructed, achieving individualized and physically feasible escape animations through a perception-decision-execution (SDM) closed loop.
ResGS: Residual Densification of 3D Gaussian for Efficient Detail Recovery
Yanzhe Lyu (University of Science and Technology of China), Xuejin Chen (University of Science and Technology of China)
RestorationData SynthesisOptimizationNeural Radiance FieldGaussian SplattingPoint Cloud
🎯 What it does: A dense operation based on Residual Split and a coarse-to-fine training pipeline (ResGS) are proposed, significantly improving detail recovery and geometric coverage in 3D Gaussian Splatting for novel view synthesis.
ResidualViT for Efficient Temporally Dense Video Encoding
Mattia Soldan (King Abdullah University of Science and Technology), Bryan Russell (Czech Institute of Informatics Robotics and Cybernetics at the Czech Technical University in Prague)
CompressionComputational EfficiencyKnowledge DistillationTransformerVideo
🎯 What it does: This paper proposes an efficient video frame feature extraction framework named ResidualViT, which reduces the computational cost of frame-level features by utilizing temporal redundancy through interleaved I-P frame encoding.
Resolving Token-Space Gradient Conflicts: Token Space Manipulation for Transformer-Based Multi-Task Learning
Wooseong Jeong (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
TransformerImage
🎯 What it does: A dynamic token modulation and expansion framework, DTME-MTL, is proposed, which utilizes the pre-trained parameters of Transformers to identify and mitigate gradient conflicts in multi-task learning, enhancing the collaborative effect between tasks.
Resonance: Learning to Predict Social-Aware Pedestrian Trajectories as Co-Vibrations
Conghao Wong (Huazhong University of Science and Technology), Beihao Xia (Huazhong University of Science and Technology)
GenerationData SynthesisTransformerGenerative Adversarial NetworkTime SeriesSequential
🎯 What it does: Proposes the Resonance (Re) model, which predicts pedestrian trajectories using the concepts of vibration and resonance, breaking down trajectory randomness into spontaneous vibration and social vibration, generated through noise sampling respectively;
ResQ: A Novel Framework to Implement Residual Neural Networks on Analog Rydberg Atom Quantum Computers
Nicholas S. DiBrita (Rice University), Tirthak Patel (Rice University)
ClassificationImageTabularPhysics RelatedOrdinary Differential Equation
🎯 What it does: The RESQ framework is proposed, implementing residual neural networks (ResNet) on both simulated and real Rydberg atom quantum computers for binary classification tasks.
Rethink Sparse Signals for Pose-guided Text-to-image Generation
Wenjie Xuan (Wuhan University), Dacheng Tao (Nanyang Technological University)
GenerationData SynthesisPose EstimationDiffusion modelImage
🎯 What it does: A text-to-image generation model based on sparse pose signals is proposed, utilizing learnable spatial pose representations and keypoint concept learning to achieve precise pose control.
Rethinking Bimanual Robotic Manipulation: Learning with Decoupled Interaction Framework
Jian-Jian Jiang (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
Robotic IntelligenceReinforcement LearningPoint CloudBenchmark
🎯 What it does: This paper proposes a decoupled interaction framework to enhance the control performance of dual-arm robots.
Rethinking Cross-Modal Interaction in Multimodal Diffusion Transformers
Zhengyao Lv (University of Hong Kong), Kwan-Yee K. Wong (University of Hong Kong)
GenerationData SynthesisTransformerSupervised Fine-TuningVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: A parameter-efficient Temperature Adaptive Cross-modal Attention (TACA) method is proposed to rebalance cross-modal attention in multi-modal diffusion transformers, thereby enhancing text-image alignment.
Rethinking Detecting Salient and Camouflaged Objects in Unconstrained Scenes
Zhangjun Zhou (Huazhong University of Science and Technology), He Tang (Huazhong University of Science and Technology)
Object DetectionSegmentationSupervised Fine-TuningImageBenchmark
🎯 What it does: This paper proposes a new framework for detecting salient and camouflaged objects in unconstrained scenarios and constructs a corresponding large-scale dataset.
Rethinking Discrete Tokens: Treating Them as Conditions for Continuous Autoregressive Image Synthesis
Peng Zheng (Jilin University), Zuxuan Wu (Fudan University)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImage
🎯 What it does: We propose the DisCon framework, which splits image generation into a two-step autoregressive process: first predicting discrete tokens and then predicting continuous tokens conditioned on the discrete tokens.
Rethinking DPO-style Diffusion Aligning Frameworks
Xun Wu (Microsoft Research), Furu Wei (Microsoft Research)
GenerationOptimizationReinforcement LearningDiffusion modelImageText
🎯 What it does: Proposes the Revised Direct Preference Optimization (RDPO) method, which aligns text-to-image diffusion models through human preferences and addresses the risks of reward estimation bias in intermediate steps and the increased probability of non-preferred samples in traditional DPO.
Rethinking Few Shot CLIP Benchmarks: A Critical Analysis in the Inductive Setting
Alexey Kravets (University of Bath), Vinay P. Namboodiri (University of Bath)
ClassificationRecognitionTransformerPrompt EngineeringContrastive LearningImageMultimodalityBenchmark
🎯 What it does: A set of unlearning techniques utilizing the CLIP model is proposed to construct a true inductive few-shot benchmark, and a self-enhanced prompt tuning method (SEPRES) is introduced based on this benchmark.
Rethinking Key-frame-based Micro-expression Recognition: A Robust and Accurate Framework Against Key-frame Errors
Zheyuan Zhang (Beijing University of Posts and Telecommunications), Hong Chen (University of Auckland)
RecognitionOptical FlowVideo
🎯 What it does: Proposes the CausalNet framework, which utilizes start-to-vertex and vertex-to-end optical flow inputs for micro-expression recognition, significantly enhancing robustness against keyframe errors.
Rethinking Layered Graphic Design Generation with a Top-Down Approach
Jingye Chen (Hong Kong University of Science and Technology), Qifeng Chen (Hong Kong University of Science and Technology)
SegmentationGenerationTransformerVision Language ModelDiffusion modelImageBenchmark
🎯 What it does: This paper proposes an Accordion framework based on a top-down process, which converts AI-generated non-hierarchical design diagrams into editable multi-layer hierarchical designs and corrects unreasonable text.
Rethinking Multi-modal Object Detection from the Perspective of Mono-Modality Feature Learning
Tianyi Zhao (Beihang University), Xingxing Wei (Beihang University)
Object DetectionKnowledge DistillationImageMultimodality
🎯 What it does: In response to the fusion degradation phenomenon in RGB-IR multimodal object detection, this paper identifies through linear probe evaluation that multimodal joint training leads to insufficient single-modal feature learning. Consequently, it proposes the M D-LIF framework: Mono-Modality Distillation (M D) enhances the capability of single-modal encoders, while Local Illumination-Aware Fusion (LIF) dynamically weights the fusion of RGB and IR features through brightness prediction.
Rethinking the Embodied Gap in Vision-and-Language Navigation: A Holistic Study of Physical and Visual Disparities
Liuyi Wang (Tongji University), Jiangmiao Pang (Shanghai AI Laboratory)
Robotic IntelligenceLarge Language ModelReinforcement LearningVision Language ModelDiffusion modelMultimodalityPoint CloudBenchmark
🎯 What it does: This paper presents VLN-PE, a Vision-and-Language Navigation platform that supports humanoid, quadruped, and wheeled robots, achieving physically realistic motion, and systematically evaluates existing self-centered VLN methods in physical environments.
Rethinking the Upsampling Process in Light Field Super-Resolution with Spatial-Epipolar Implicit Image Function
Ruixuan Cong (Beihang University), Hao Sheng (Beihang University)
RestorationSuper ResolutionImage
🎯 What it does: A light field super-resolution decoder based on implicit image functions (SEIIF) is proposed, achieving arbitrary scale light field super-resolution through two upsampling modes: spatial (SIIF) and epipolar (EIIF).
Retinex-MEF: Retinex-based Glare Effects Aware Unsupervised Multi-Exposure Image Fusion
Haowen Bai (Xi'an Jiaotong University), Shuang Xu (Northwestern Polytechnical University)
RestorationConvolutional Neural NetworkImage
🎯 What it does: An unsupervised Retinex-MEF method is proposed, which utilizes Retinex decomposition to separate multi-exposure images into shared reflectance and individual illumination, explicitly modeling the glare caused by overexposure, thus achieving controllable fusion of exposures.
RetinexMCNet: A Memory Controller Dominated Network for Low-Light Video Enhancement Based on Retinex
Meiao Wang (Beijing University of Posts and Telecommunications), Jie Xu (Tsinghua University)
RestorationConvolutional Neural NetworkVideo
🎯 What it does: A two-stage low-light video enhancement network, RetinexMCNet, is proposed, which first performs single-frame denoising and exposure correction based on Retinex theory, and then achieves global temporal consistency by fusing historical key frames through a memory controller (MC).
ReTracker: Exploring Image Matching for Robust Online Any Point Tracking
Dongli Tan (Zhejiang University), Xiaowei Zhou (Zhejiang University)
Object TrackingTransformerVideo
🎯 What it does: This paper proposes an online point tracking framework called ReTracker, which is capable of re-tracking points after long-term occlusion and processing video in real-time.
Reusing Computation in Text-to-Image Diffusion for Efficient Generation of Image Sets
Dale Decatur (University of Chicago), Matheus Gadelha (Adobe Research)
GenerationComputational EfficiencyDiffusion modelImageText
🎯 What it does: This study investigates a training-independent, automated hierarchical clustering method that shares early denoising computations in text-to-image diffusion models, significantly reducing computational consumption when generating multiple images.
Revelio: Interpreting and leveraging semantic information in diffusion models
Dahye Kim (Boston University), Deepti Ghadiyaram (Runway)
ClassificationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningConvolutional Neural NetworkDiffusion modelAuto EncoderImage
🎯 What it does: This paper provides a mechanistic explanation and validation of the visual semantic information at different layers and time steps of diffusion models by training a k-sparse autoencoder and a lightweight classifier.
Reverse Convolution and Its Applications to Image Restoration
Xuhong Huang (Nanjing University), Lei Zhang (Hong Kong Polytechnic University)
RestorationSuper ResolutionConvolutional Neural NetworkTransformerImage
🎯 What it does: A deep convolutional reverse convolution operation is proposed, along with the construction of a corresponding network module, applied to image denoising, super-resolution, and deblurring tasks.
Revisiting Adversarial Patch Defenses on Object Detectors: Unified Evaluation, Large-Scale Dataset, and New Insights
Junhao Zheng (Xi'an Jiaotong University), Chao Shen (Xi'an Jiaotong University)
Object DetectionAdversarial AttackDiffusion modelImageBenchmark
🎯 What it does: The first benchmark for adversarial patch defense against object detection (APDE) is proposed, systematically evaluating 11 defense methods, 13 patch attacks, and 11 detectors, and constructing a multi-type patch dataset of 94,000 images.
Revisiting Efficient Semantic Segmentation: Learning Offsets for Better Spatial and Class Feature Alignment
Shi-Chen Zhang (Nankai University), Ming-Ming Cheng (Nankai University)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a new dual-branch offset learning paradigm to simultaneously adjust image features and class representations, thereby addressing the feature-class misalignment issue in traditional pixel-level classification methods for lightweight semantic segmentation.
Revisiting Image Fusion for Multi-Illuminant White-Balance Correction
David Serrano-Lozano (Universitat Autònoma de Barcelona), Javier Vazquez-Corral (Computer Vision Center)
RestorationTransformerImage
🎯 What it does: This study investigates white balance correction in multi-lighting scenarios and proposes a method for nonlinear fusion of multiple preset WB images based on Transformer, while constructing a large-scale multi-lighting sRGB dataset.
Revisiting Point Cloud Completion: Are We Ready For The Real-World?
Stuti Pathak (University of Antwerp), Rudi Penne (University of Antwerp)
RestorationData SynthesisTransformerPoint CloudBenchmark
🎯 What it does: This paper presents research on point cloud completion for real-world applications, constructing the RealPC dataset, which contains approximately 40,000 pairs and 21 categories of industrial structures, and explores the role of 0-dimensional persistent homology (PH) in point cloud completion.
Revisiting Pool-based Prompt Learning for Few-shot Class-incremental Learning
Yongwei Jiang (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
ClassificationRecognitionTransformerPrompt EngineeringImage
🎯 What it does: This paper re-examines pool-based prompt learning in few-shot class-incremental learning (FSCIL) and finds that traditional stacking prompts in the token dimension leads to performance drops for new classes due to information saturation. It proposes LGSP-Prompt, which reconstructs prompts in the spatial dimension through local and global spatial prompts to address the saturation issue.
RGE-GS: Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors
Sicong Du (CaiNiao Inc. Alibaba Group), Sheng Yang (CaiNiao Inc. Alibaba Group)
RestorationGenerationAutonomous DrivingDiffusion modelGaussian SplattingImagePoint Cloud
🎯 What it does: A reward-guided Gaussian Splatting extended reconstruction framework (RGE-GS) is proposed, which can utilize prior images generated by diffusion models to complete missing road scenes in a single scan.
RhythmGuassian: Repurposing Generalizable Gaussian Model For Remote Physiological Measurement
Hao Lu (Hong Kong University of Science and Technology), Yingcong Chen (Hong Kong University of Science and Technology)
RecognitionData SynthesisGaussian SplattingVideo
🎯 What it does: The RhythmGaussian method is proposed, utilizing the Generalizable Gaussian Model (GGM) to achieve geometric and chromatic decoupling of 4D Gaussian representations for high-quality extraction of remote photoplethysmography (rPPG) signals, significantly reducing interference from motion and lighting noise.
RI3D: Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors
Avinash Paliwal (Texas A&M University), Nima Khademi Kalantari (Texas A&M University)
RestorationGenerationData SynthesisDiffusion modelGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes a sparse view synthesis method based on 3D Gaussian splatting, utilizing two personalized diffusion models (repair and inpainting) to reconstruct visible areas and fill in missing areas in a two-stage optimization process.
Riemannian-Geometric Fingerprints of Generative Models
Hae Jin Song (University of Southern California), Laurent Itti (University of Southern California)
GenerationData SynthesisOptimizationDiffusion modelGenerative Adversarial NetworkImageMultimodality
🎯 What it does: A definition of generative model fingerprint based on Riemannian geometry is proposed, and a gradient algorithm for calculating fingerprints from finite samples is provided, thereby achieving model attribution.
RIOcc: Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction
Baojie Fan (Nanjing University of Posts and Telecommunications), Huijie Fan (Shenyang Institute of Automation, Chinese Academy of Sciences)
SegmentationAutonomous DrivingComputational EfficiencyTransformerSupervised Fine-TuningMultimodalityPoint Cloud
🎯 What it does: The RIOcc framework is proposed, which combines LiDAR and camera information for 3D semantic occupancy prediction in a unified BEV space, significantly reducing computational costs and improving accuracy.
RIPE: Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction
Johannes Künzel (Humboldt University Berlin), Peter Eisert (Fraunhofer Heinrich Hertz Institute)
Reinforcement LearningImage
🎯 What it does: Proposes the RIPE framework, which utilizes weakly supervised image pairs to train robust keypoint detection and description.
RMultiplex200K: Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications
Sijia Chen (Hong Kong University of Science and Technology), Bin Song (Xidian University)
TransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: The first multimodal dataset for the telecommunications field, RMultiplex200K, has been constructed, and an automated planning-based annotation process (ApPA) without manual labeling has been used to generate reasoning processes with step-level correctness scores; the TC-NAVIGATOR multimodal process reward model has been proposed to validate and enhance the step-by-step reasoning of VLM.