CVPR 2026 Papers — Page 28
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
Prototype-Guided Concept Erasure in Diffusion Models
Yuze Cai (Fudan University), Hong Lu (Fudan University)
GenerationTransformerPrompt EngineeringDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: Proposes a training-free concept erasure method based on prototype guidance, which can effectively eliminate general concepts during diffusion model inference.
Prototypical Action Reasoning Facilitated by Vision-Language Alignment for Egocentric Action Anticipation
Jiang Shao (Northwestern Polytechnical University), Xiaochun Zou (Northwestern Polytechnical University)
RecognitionTransformerVision Language ModelVision-Language-Action ModelVideoTextMultimodality
🎯 What it does: This paper proposes a prototype action reasoning framework based on vision-language alignment (PAR-VLA) for front-view perspective action prediction.
Proxy-GS: Unified Occlusion Priors for Training and Inference in Structured 3D Gaussian Splatting
Yuanyuan Gao (Shanghai Artificial Intelligence Laboratory), Zhihang Zhong (Shanghai Jiao Tong University)
Computational EfficiencyGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes Proxy-GS, which utilizes a lightweight proxy grid to realize an occlusion-aware Gaussian Splatting training and inference framework.
Proxy-Tuning: Tailoring Multimodal Autoregressive Models for Subject-Driven Image Generation
Yi Wu, Bin Li (University Of Hong Kong)
GenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelImageTextMultimodality
🎯 What it does: Studies how to leverage diffusion model supervision to fine-tune autoregressive image generation models for high-quality subject-driven image generation
Proxy3D: Efficient 3D Representations for Vision-Language Models via Semantic Clustering and Alignment
Jerry Jiang (Tsinghua University), Wenzhao Zheng (UC Berkeley)
Computational EfficiencyRepresentation LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningVision Language ModelVision-Language-Action ModelVideoMultimodality
🎯 What it does: Propose the Proxy3D method, which extracts semantic and geometric features from video frames, generates compact 3D proxy representations through semantic clustering, and combines with VLM under multi-stage training to achieve efficient spatial reasoning.
ProxyFL: A Proxy-Guided Framework for Federated Semi-Supervised Learning
Duowen Chen (East China Normal University), Yan Wang (East China Normal University)
ClassificationFederated LearningContrastive LearningImage
🎯 What it does: In the federated semi-supervised learning framework, we propose using classifier learnable weights as a unified 'proxy' to simultaneously alleviate inter-client (external) and label-unlabeled data mismatch (internal) heterogeneity, and enhance model performance through two modules: Global Proxy Tuning (GPT) and Uncertain Class Proxy Learning (ICPL).
PRUE: A Practical Recipe for Field Boundary Segmentation at Scale
Gedeon Muhawenayo (Arizona State University), Hannah Kerner (Arizona State University)
SegmentationConvolutional Neural NetworkImageAgriculture Related
🎯 What it does: This paper proposes a PRUE method for large-scale, cross-regional agricultural field boundary segmentation. By combining the U-Net architecture, EfficientNet-B7 encoder, log-cosh Dice loss, boundary weighting, brightness/scale augmentation, and channel shuffling techniques, it achieves robust segmentation of Sentinel-2 L2A images and generates high-precision field boundaries on the FTW benchmark.
Prune Wisely, Reconstruct Sharply: Compact 3D Gaussian Splatting via Adaptive Pruning and Difference-of-Gaussian Primitives
Haoran Wang (University of Bristol), Nantheera Anantrasirichai
Computational EfficiencyGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes an adaptive reconstruction-aware pruning scheduler (RPS) and a 3D differential Gaussian (3D-DoG) primitive capable of simultaneously representing positive and negative densities, achieving significant model compression and visual quality improvement under the 3D Gaussian splatting framework.
Prune2Drive: A Plug-and-Play Framework for Accelerating Vision-Language Models in Autonomous Driving
Minhao Xiong (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
Autonomous DrivingOptimizationComputational EfficiencyHyperparameter SearchVision Language ModelMultimodality
🎯 What it does: Propose Prune2Drive, a plug-and-play visual token pruning framework designed to accelerate the inference of multi-view vision-language models in autonomous driving.
PS-SR: Pseudo-Single-Step Video Super-Resolution via Speculative Diffusion
Aiqiu Wu (University of Science and Technology of China), Tao Mei (HiDream.ai Inc)
Super ResolutionTransformerDiffusion modelVideo
🎯 What it does: Propose a pseudo single-step video super-resolution framework PS-SR, which first uses a powerful base model to sample globally once to obtain the overall structure, then employs a lightweight draft model and frequency domain update rules for multi-step refinement, adding details only in the high-frequency domain to balance speed and quality.
PSDesigner: Automated Graphic Design with a Human-Like Creative Workflow
Xincheng Shuai (Fudan University), Dacheng Tao (Nanyang Technological University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelMultimodality
🎯 What it does: Proposed PSDesigner, a graphics design system capable of automatically generating editable PSD files based on user instructions, achieving human-like designer workflows through asset collection, graphic planning, and tool execution.
PSR: Scaling Multi-Subject Personalized Image Generation with Pairwise Subject-Consistency Rewards
Shulei Wang (Zhejiang University), Qi Tian (Huawei Inc)
GenerationData SynthesisTransformerSupervised Fine-TuningReinforcement LearningFlow-based ModelImageText
🎯 What it does: Proposed a scalable multi-agent personalized image generation framework, leveraging a single-agent generative model to construct 350K high-quality multi-agent training data, and subsequently employing frame-level position encoding and Pairwise Subject-Consistency Rewards for post-training reinforcement learning.
PTC-Depth: Pose-Refined Monocular Depth Estimation with Temporal Consistency
Leezy Han (Ajou University), Hyeonbeom Lee (Ajou University)
Depth EstimationOptical FlowImagePoint Cloud
🎯 What it does: Integrate wheel odometry with a monocular camera, estimate camera pose using optical flow, and obtain sparse depth via triangulation; then fuse relative depth (from a base model) with sparse depth through recursive Bayesian fusion to achieve temporally consistent and metric-scale dense depth maps.
PureCC: Pure Learning for Text-to-Image Concept Customization
Zhichao Liao (Tsinghua University), Liang Pan (Nanyang Technological University)
GenerationSupervised Fine-TuningFlow-based ModelImage
🎯 What it does: Propose the PureCC method for text-to-image concept customization, aiming to maintain the original model's behavior and capabilities when introducing new concepts.
PureProof: Diffusion-Resistant Black-box Targeted Attack on Large Vision-Language Models
Yiming Cao (Hong Kong Polytechnic University), Bin Xiao (Hong Kong Polytechnic University)
Adversarial AttackVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Propose a black-box targeted adversarial attack framework called PureProof for large vision-language models (VLMs), which can successfully induce the model to output predefined text under Diffusion-Based Purification (DBP) defense;
Push-and-Step: From RL-Based Balance Recovery to Physical Simulation of Dense Crowds
Alexis Jensen (Univ Rennes, Inria, CNRS, IRISA-UMR 6074), Julien Pettré (Univ Rennes, Inria, CNRS, IRISA-UMR 6074)
Reinforcement LearningVideo
🎯 What it does: This study develops a physics-based full-body agent simulation method that enables agents to recover balance in high-density crowds through gait or hand contact, and achieve adaptive responses to external impacts via reinforcement learning.
Pushing the Frontier of Audiovisual Perception with Large-Scale Multimodal Correspondence Learning
Apoorv Vyas (Meta Superintelligence Labs), Wei-Ning Hsu (Meta Superintelligence Labs)
ClassificationData SynthesisRetrievalRepresentation LearningTransformerContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: Proposes Perception Encoder Audiovisual (PEAV), a unified multi-modal encoder capable of simultaneously aligning and embedding audio, video, and text, achieving state-of-the-art performance across multiple tasks such as audio, video, music, and speech under zero-shot settings.
PV-Ground: Text-Guided Point-Voxel Interaction for 3D Visual Grounding
Junpeng Shang (Zhejiang University), Dongfang Ma (Zhejiang University)
Object DetectionSegmentationConvolutional Neural NetworkTransformerVision Language ModelTextPoint Cloud
🎯 What it does: Propose a text-guided point-voxel interaction framework, PV-Ground, for 3D visual localization tasks, which achieves efficient fusion of text and point cloud features while preserving high-resolution geometric details.
PvP: Data-Efficient Humanoid Robot Learning with Proprioceptive-Privileged Contrastive Representations
Mingqi Yuan (HK PolyU), Wenjun Zeng (EIT)
Robotic IntelligenceReinforcement LearningAuto EncoderContrastive LearningTabularTime Series
🎯 What it does: In this study, the authors propose the PvP framework, which enhances representation learning and policy training in whole-body control (WBC) tasks through contrastive learning between self-sensory states and privileged states.
PyramidalWan: On Making Pretrained Video Model Pyramidal for Efficient Inference
Denis Korzhenkov (Qualcomm AI Research), Amirhossein Habibian (Qualcomm AI Research)
GenerationComputational EfficiencyKnowledge DistillationSupervised Fine-TuningDiffusion modelFlow-based ModelAuto EncoderVideoTextBenchmark
🎯 What it does: Convert existing large-scale pre-trained video diffusion models into a pyramid structure, and perform distillation training with multi-step and few-step inference based on this, significantly reducing inference costs without compromising visual quality.
PyraTok: Language-Aligned Pyramidal Tokenizer for Video Understanding and Generation
Onkar Susladkar (University Of Illinois Urbana Champaign), Ismini Lourentzou (University Of Illinois Urbana Champaign)
ClassificationRestorationSegmentationGenerationVision Language ModelAuto EncoderVideoTextMultimodality
🎯 What it does: Constructed a pyramid-style video tokenizer called PyraTok, which performs discrete quantization of video features at multiple scales and achieves tight coupling between text and vision through dual semantic alignment, thereby enhancing the performance of video reconstruction, text-to-video generation, and multimodal understanding tasks.
QD-PCQA: Quality-Aware Domain Adaptation for Point Cloud Quality Assessment
Guohua Zhang (Beijing Jiaotong University), Weisi Lin (Nanyang Technological University)
Domain AdaptationConvolutional Neural NetworkGenerative Adversarial NetworkImagePoint Cloud
🎯 What it does: Perform unsupervised domain adaptation between source domain images and target domain point clouds to achieve no-reference point cloud quality assessment (NR-PCQA)
QuadSync: Quadrifocal Tensor Synchronization via Tucker Decomposition
Daniel Miao (University of Minnesota), Joe Kileel (University of Texas at Austin)
Pose EstimationOptimizationImage
🎯 What it does: This paper studies the multi-view synchronization problem, proposes the Block Quadrifocal Tensor and designs a global synchronization algorithm called QuadSync based on its Tucker decomposition, as well as a Joint Opt. method for jointly synchronizing trifocal, bifocal, and quadrifocal tensors.
Quant Experts: Token-aware Adaptive Error Reconstruction with Mixture of Experts for Large Vision-Language Models Quantization
Chenwei Jia (Xi'an Jiaotong University), Hongbin Sun (Xi'an Jiaotong University)
Computational EfficiencyTransformerMixture of ExpertsMultimodality
🎯 What it does: Propose a Token-aware Adaptive Error Reconstruction method (Quant Experts, QE) for post-training quantization (PTQ) in Vision-Language Models, by separating token-independent and token-dependent important channels, and compensating quantization errors using shared low-rank adapters and routing experts respectively.
QUANTIPHY: A Quantitative Benchmark Evaluating Physical Reasoning Abilities of Vision-Language Models
Li Puyin (Stanford University), Ehsan Adeli (Stanford University)
TransformerPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmarkPhysics RelatedChain-of-Thought
🎯 What it does: Constructed the QUANTIPHY benchmark to assess the quantitative reasoning capabilities of VLMs in physics kinematics through videos and numerical priors.
Quantized Residuals to Continuous Prompts for Few-Shot Class Incremental Learning in Vision-Language Models
Abhishek Kumar Sinha (Indian Institute of Science), Soma Biswas (Indian Institute of Science)
ClassificationRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose a QR-Prompt framework that leverages visual-textual residual quantization to map into continuous prompts, achieving Few-Shot Class-Incremental Learning (FSCIL) on existing Vision-Language Models (VLMs).
Quantum-Gated Task-interaction Knowledge Distillation for Pre-trained Model-based Class-Incremental Learning
Linjie Li (Beijing University of Posts and Telecommunications), Yang Ji (Beijing University of Posts and Telecommunications)
ClassificationKnowledge DistillationImage
🎯 What it does: Studying class-incremental learning based on pre-trained models, proposing a task-interaction knowledge distillation (QKD) framework based on quantum gates;
QuantVLA: Scale-Calibrated Post-Training Quantization for Vision-Language-Action Models
Jingxuan Zhang (Ohio State University), Mi Zhang (Ohio State University)
Computational EfficiencyRobotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelMultimodality
🎯 What it does: This paper proposes QuantVLA, a training-free post-training quantization framework designed to low-precisionize the language encoder and diffusion transformer (DiT) action head in vision-language-action (VLA) models.
QuCNet: Quantum Deep Learning Driven Multi-Circuit Network for Remote Sensing Image Classification
Komal Komal, Subrahmanyam Murala (Trinity College Dublin)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: Proposed a hybrid quantum-classical network called QuCNet, combining a lightweight convolutional encoder and 16 four-qubit trainable quantum circuits for remote sensing image classification.
Query2Uncertainty: Robust Uncertainty Quantification and Calibration for 3D Object Detection under Distribution Shift
Till Beemelmanns (RWTH Aachen), Lutz Eckstein (RWTH Aachen)
Object DetectionTransformerFlow-based ModelPoint Cloud
🎯 What it does: Propose a post-calibration method called Query2Uncertainty based on query feature density, which uses the object query features from DETR-style 3D object detectors (e.g., PETR, SECOND) to estimate the distribution density of input samples. The method dynamically adjusts classification confidence and regression uncertainty based on the density, making the model's confidence more reliable under distribution drift scenarios (e.g., rainy days, snowy days, noise).
QueryMe: Query-Driven Open-Vocabulary 3D Object Affordances Grounding from Multimodal Evidence
Weiyu Zhao (Harbin Institute of Technology), Shengping Zhang (Harbin Institute of Technology)
SegmentationConvolutional Neural NetworkTransformerLarge Language ModelVision-Language-Action ModelImageTextMultimodalityPoint Cloud
🎯 What it does: Propose the QueryMe framework, which utilizes multimodal information from human-object interaction (HOI) images, text descriptions, and 3D point clouds to perform open-vocabulary functional region localization on 3D objects.
QueryOcc: Query-based Self-Supervision for 3D Semantic Occupancy
Adam Lilja (Chalmers University of Technology), Lars Hammarstrand (Chalmers University of Technology)
Autonomous DrivingRepresentation LearningGaussian SplattingImagePoint CloudBenchmark
🎯 What it does: Learn a continuous 3D semantic occupancy field directly from multi-view images using a query-based self-supervised method.
Question-guided Visual Compression with Memory Feedback for Long-Term Video Understanding
Sosuke Yamao (Fujitsu Research), Shun Takeuchi (Fujitsu Research)
CompressionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningVideoMultimodality
🎯 What it does: Proposes the QViC-MF framework, which utilizes problem-guided multi-frame visual compression and memory feedback to achieve efficient understanding of long videos.
QuietPrune: Query-Guided Early Token Pruning for Vision-Language Models
Tianxiao Gao (Ant Group), Chenguang Ma (Ant Group)
Computational EfficiencyTransformerVision Language ModelMultimodalityBenchmark
🎯 What it does: In visual-language models, this paper proposes the QuietPrune method, which performs query-guided early visual token pruning within ViT to reduce computational costs.
Quota-Calibrated Fine-Grained Alignment with Context-Aware Marginals for Text-based Person Retrieval
Dongsheng Li (Northeast Normal University), Qiushi Xia (Northeast Normal University)
RetrievalVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes QC-Align, a fine-grained alignment framework for text-based person retrieval (TPR), aiming to address the issues of traditional methods ignoring matching capacity and over-concentrating weights.
QVGGT: Post-Training Quantized Visual Geometry Grounded Transformer
Zhizhen Pan (Westlake University), Huan Wang (Westlake University)
Pose EstimationDepth EstimationCompressionComputational EfficiencyTransformerImagePoint Cloud
🎯 What it does: Post-training quantization of the large-scale visual geometry Transformer VGGT, proposing the QVGGT framework to achieve low-bit weight quantization while maintaining 3D reconstruction accuracy.
Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition
Shengming Yin (Hong Kong University of Science and Technology (Guangzhou)), Chenfei Wu (Alibaba)
GenerationTransformerDiffusion modelAuto EncoderImageTextMultimodality
🎯 What it does: Propose Qwen-Image-Layered, an end-to-end diffusion model that can decompose a single RGB image into multiple layers of semantically disentangled RGBA layers, achieving naturally editable image representations.
R-4B: Incentivizing General-Purpose Auto-Thinking in MLLMs via Bi-Mode Annealing and Reinforce Learning
Qi Yang (University of Chinese Academy of Sciences), Houwen Peng (Tencent HunYuan)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Designed and implemented R-4B, a multimodal large language model with two output modes (thinking and direct answering), and enabled the model to adaptively decide whether to perform chain-of-thought reasoning during inference through bi-mode annealing training (Bi-mode Annealing) and bi-mode policy optimization (BPO).
R-C2: Cycle-Consistent Reinforcement Learning Improves Multimodal Reasoning
Zirui Zhang (Rutgers University), Chengzhi Mao (Rutgers University)
TransformerReinforcement LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper proposes a cross-modal cyclic consistency reinforcement learning framework, R-C², to address the issue of answer inconsistency between visual and text modalities in multi-modal large language models.
R$^2$TUA: Reconstruction-residual Based Targeted and Untargeted Attack Against Text-Image Person Re-Identification
Yubo Wang (University of Science and Technology of China), Jixiang Niu (University of Science and Technology of China)
RetrievalAdversarial AttackTransformerContrastive LearningMultimodality
🎯 What it does: A two-stage adversarial attack framework named R²TUA is designed for text-image person re-identification (TI-ReID) models, capable of generating fine-grained and imperceptible perturbations while achieving both targeted and untargeted attacks.
R2-Seg: Training-Free OOD Medical Tumor Segmentation via Anatomical Reasoning and Statistical Rejection
Shuaike Shen (Carnegie Mellon University), Shangqi Gao (Zhejiang University)
SegmentationLarge Language ModelBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: R-Seg2 proposes a training-agnostic OOD tumor segmentation framework, first leveraging LLM for anatomical reasoning to plan ROI, then performing segmentation within these ROI using frozen BiomedParse, and filtering false positives through two-sample statistical tests.
R2G: A Multi-View Circuit Graph Benchmark Suite from RTL to GDSII
Zewei Zhou (Nanjing University of Science and Technology), Daying Sun (Nanjing University of Science and Technology)
Graph Neural NetworkGraphBenchmark
🎯 What it does: Proposed the R2G multi-view circuit diagram benchmark, providing a complete DEF-to-Graph conversion pipeline from RTL to GDSII, and constructed a circuit diagram dataset containing five information-aligned views.
R3-PCQA: Ray-Reprojection-Reinforcement for No-Reference 3D Point Cloud Quality Assessment
Junhyuk Seo (Hansung University), Heeseok Oh (Hansung University)
Convolutional Neural NetworkTransformerReinforcement LearningPoint Cloud
🎯 What it does: Proposed a novel no-reference 3D point cloud quality assessment framework, R3-PCQA, which utilizes geometry-aware ray projection to achieve 2D-3D spatial correspondence, employs a reinforcement learning-based quality significant subcloud selector (QSS) to dynamically select local subclouds most influential for quality assessment, and integrates multi-view features into a global quality score through a global perspective attention module.
R4-CGQA: Retrieval-based Vision Language Models for Computer Graphics Image Quality Assessment
Zhuangzi Li (Nanyang Technological University), Weisi Lin (Nanyang Technological University)
RetrievalVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposed a retrieval-enhanced two-stream framework named R4-CGQA to improve the ability of large models in computer graphics image quality assessment (CGQA), and constructed a dataset containing 3.5K CG images and six-dimensional quality descriptions.
R4: Retrieval-Augmented Reasoning for Vision-Language Models in 4D Spatio-Temporal Space
Tin Stribor Sohn (Dr. Ing. h.c. F. Porsche AG), Eric Sax (Karlsruhe Institute of Technology)
RetrievalVision Language ModelSimultaneous Localization and MappingVideoTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Developed a training-free retrieval-enhanced 4D spatiotemporal memory framework called R4, enabling vision-language models (VLMs) to answer long-span spatiotemporal questions by retrieving and reasoning over continuous 4D memories.
R4Det: 4D Radar-Camera Fusion for High-Performance 3D Object Detection
Zhongyu Xia (Peking University), Weijun Qin (EBTech Co. Ltd)
Object DetectionAutonomous DrivingConvolutional Neural NetworkRecurrent Neural NetworkGaussian SplattingImageMultimodalityPoint Cloud
🎯 What it does: Developed the R4Det framework to achieve high-performance 3D object detection with 4D radar-camera fusion.
RAAS: LLM Agentic System Architecture Search with GRPO
Jiayi Yang (Wuhan University), Mang Ye (Wuhan University)
Neural Architecture SearchLarge Language ModelAgentic AIText
🎯 What it does: Proposed a robust architecture-adaptive search framework RAAS for automatically discovering high-quality multi-agent workflows in Agentic Supernet;
Radar-Guided Polynomial Fitting for Metric Depth Estimation
Patrick Rim (Yale University), Alex Wong (Yale University)
Depth EstimationAutonomous DrivingConvolutional Neural NetworkImagePoint Cloud
🎯 What it does: A scale-free depth map is obtained using a pre-trained monocular depth estimation model, followed by using radar point cloud information to guide polynomial fitting, ultimately generating an accurate metric depth map.
RADAR: VQ-VAE Decoder of VAR is a Good Student for Restoring Against Degradation by Acceleration
Ziyang Wang (Beijing University of Posts and Telecommunications), Xueming Li (Beijing University of Posts and Telecommunications)
GenerationComputational EfficiencyKnowledge DistillationAuto EncoderImageTextMultimodality
🎯 What it does: Proposes a two-stage acceleration framework RADAR, including semantic cost-aware mask SCA-Mask and post-acceleration adaptation PAA, which can significantly improve the inference speed of VAR models while maintaining image quality.
Radiance Meshes for Volumetric Reconstruction
Alexander Mai (University of California, San Diego), Jonathan T. Barron (Google)
Neural Radiance FieldPoint CloudMesh
🎯 What it does: Proposes a radiance field representation called Radiance Meshes based on Delaunay tetrahedral meshes, where each tetrahedron has constant density and linear color, supporting hardware rasterization and ray tracing to achieve high-quality real-time view synthesis.
RAG-TP: A General Framework for Vehicle Trajectory Prediction via Retrieval-Augmented Generation
Ziyi Wang (National University of Defense Technology), Shaowu Yang (National University of Defense Technology)
Autonomous DrivingMixture of ExpertsTime SeriesRetrieval-Augmented Generation
🎯 What it does: Propose a retrieval-augmented generation framework RAG-TP, which reformulates the vehicle trajectory prediction task as a process of retrieving prior knowledge from a large-scale knowledge base, decoupling real-time perception from offline experience.
RaGS: Unleashing 3D Gaussian Splatting from 4D Radar and Monocular Cue for 3D Object Detection
Xiaokai Bai, Hui-Liang Shen (Zhejiang University)
Object DetectionGaussian SplattingMultimodalityBenchmark
🎯 What it does: Proposes the RaGS framework, which utilizes 3D Gaussian splats to fuse 4D mmWave radar with monocular images, achieving sparse and efficient 3D object detection.
RAGTrack: Language-aware RGBT Tracking with Retrieval-Augmented Generation
Hao Li (Army Engineering University of PLA), Huchuan Lu (Dalian University of Technology)
Object TrackingTransformerLarge Language ModelVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposes the RAGTrack framework, which utilizes a multi-modal Transformer Encoder to fuse RGB, TIR, and text information, and achieves language-driven RGBT tracking through Adaptive Token Fusion (ATF) and Context-aware Reasoning Module (CRM);
RAID: Retrieval-Augmented Anomaly Detection
Mingxiu Cai (Northeastern University), Xiatian Zhu (University of Surrey)
Anomaly DetectionTransformerMixture of ExpertsImageRetrieval-Augmented Generation
🎯 What it does: Propose an unsupervised anomaly detection method RAID based on the retrieval-augmented generation (RAG) framework, which uses retrieved normal samples to guide the generation of fine-grained anomaly maps.
RAISE: Requirement-Adaptive Evolutionary Refinement for Training-Free Text-to-Image Alignment
Liyao Jiang (University of Alberta), Di Niu (University of Alberta)
GenerationTransformerLarge Language ModelVision Language ModelDiffusion modelImageText
🎯 What it does: Propose a training-free, demand-driven adaptive evolution framework called RAISE, which gradually improves the alignment quality between text and images during the inference phase through multi-round evolution and tool validation.
RAM: Recover Any 3D Human Motion in-the-Wild
Sen Jia (University of Washington), Lei Li (Beijing Institute of Technology)
Pose EstimationTransformerVideoMesh
🎯 What it does: Proposed a real-time multi-person 3D human motion recovery framework named RAM, integrating motion-aware tracking, spatiotemporal memory regression, motion prediction, and fusion modules to continuously retrieve SMPL meshes from monocular videos;
RAMEN: Resolution-Adjustable Multimodal Encoder for Earth Observation
Nicolas Houdré (University of Paris Cité), Sylvain Lobry (University of Paris Cité)
SegmentationComputational EfficiencyRepresentation LearningData-Centric LearningTransformerMixture of ExpertsAuto EncoderMultimodalityTime SeriesBenchmark
🎯 What it does: Proposes RAMEN—a tunable resolution multimodal encoder for unified representation learning of Earth observation data, achieving cross-modal transfer through self-supervised mask reconstruction.
Random Wins All: Rethinking Grouping Strategies for Vision Tokens
Qihang Fan (MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences), Ran He (MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences)
ClassificationObject DetectionSegmentationTransformerImageMultimodalityPoint Cloud
🎯 What it does: Proposed an extremely simple random grouping strategy as an alternative to the complex grouping methods in Vision Transformers, and validated its effectiveness across various visual tasks (image classification, object detection, instance segmentation, semantic segmentation, point cloud segmentation, and vision-language tasks).
Rank-Guided Pseudo-Bias Learning for Robust Black-Box Adaptation
Rajeev Ranjan Dwivedi (IISER Bhopal), Vinod K Kurmi (IISER Bhopal)
ClassificationDomain AdaptationContrastive LearningImage
🎯 What it does: This paper proposes a full black-box debiasing framework called PLD-Debias, which performs fair and robust classification using a frozen pre-trained visual encoder;
RankOOD - Class Ranking-based Out-of-Distribution Detection
Dishanika Denipitiyage (University of Sydney), Sanjay Chawla (Qatar Computing Research Institute)
Anomaly DetectionImage
🎯 What it does: Proposed a category ranking-based OOD detection framework called RankOOD, which leverages category ranking information from pre-trained models for training and inference.
RAP: Fast Feedforward Rendering-Free Attribute-Guided Primitive Importance Score Prediction for Efficient 3D Gaussian Splatting Processing
Kaifa Yang (Shanghai Jiao Tong University), Zhu Li (University of Missouri-Kansas City)
OptimizationComputational EfficiencyGaussian SplattingPoint Cloud
🎯 What it does: Propose RAP, an attribute-guided, rendering-free, efficient 3D Gaussian profile importance score prediction method.
RaPA: Enhancing Transferable Targeted Attacks via Random Parameter Pruning
Tongrui Su (Institute of Computing Technology, Chinese Academy of Sciences), Xueqi Cheng (Institute of Computing Technology, Chinese Academy of Sciences)
Adversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: Studied a Random Parameter Pruning Attack (RaPA), which enhances the success rate of transferable targeted attacks by randomly pruning parameters of the substitute model during the attack process.
RAPID: Reusing Attention Sparsity with Inter-step Adaptation for Efficient Video Diffusion
Shangran Lin (Alibaba Cloud Computing), Qiang Liu (Alibaba Cloud Computing)
GenerationComputational EfficiencyTransformerDiffusion modelVideo
🎯 What it does: Propose the RAPID framework, which significantly improves the generation speed and quality of video diffusion models by calculating attention importance once and caching it, while reusing sparse attention masks.
RARE: Learn to RAnk and REtrieve for Monocular 3D Object Detection
Hyeonjeong Park (University of Illinois Chicago), Wei Tang (University of Illinois Chicago)
Object DetectionRetrievalAutonomous DrivingTransformerImage
🎯 What it does: Propose a unified Rank-and-Retrieve framework for monocular 3D object detection, improving confidence estimation and achieving multi-modal 3D localization.
Rascene: High-Fidelity 3D Scene Imaging with mmWave Communication Signals
Kunzhe Song (Michigan State University), Huacheng Zeng (Michigan State University)
GenerationDepth EstimationConvolutional Neural NetworkMultimodalityPoint Cloud
🎯 What it does: Utilize millimeter-wave OFDM communication signals to achieve high-fidelity 3D scene imaging on a single device, proposing a multi-frame confidence-weighted fusion framework to overcome the sparsity and multipath interference of RF signals;
Rationale-Enhanced Decoding for Multi-modal Chain-of-Thought
Shin'ya Yamaguchi (NTT), Daiki Chijiwa (NTT)
Large Language ModelVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose a no-training Rationale-Enhanced Decoding (RED) that better leverages intermediate reasoning steps (rationale) for final answer generation by combining the probability distributions of visual and reasoning steps in multi-modal chain-of-thought (CoT) reasoning.
RaUF: Learning the Spatial Uncertainty Field of Radar
Shengpeng Wang (Huazhong University of Science and Technology), Wei Wang (Wuhan University)
Autonomous DrivingConvolutional Neural NetworkPoint Cloud
🎯 What it does: Propose the RaUF framework, achieving high-confidence dense radar point cloud reconstruction by learning the spatial uncertainty field of radar and utilizing Doppler consistency.
RAVEN: Erasing Invisible Watermarks via Novel View Synthesis
Fahad Shamshad (MBZUAI), Karthik Nandakumar (MBZUAI)
GenerationAdversarial AttackDiffusion modelImage
🎯 What it does: The study proposes a view synthesis attack based on a zero-shot diffusion model, capable of erasing invisible watermarks without accessing watermark information or models.
RAVEN: Radar Adaptive Vision Encoders for Efficient Chirp-wise Object Detection and Segmentation
Anuvab Sen (Georgia Institute of Technology), Saibal Mukhopadhyay (Georgia Institute of Technology)
Object DetectionSegmentationAutonomous DrivingTransformerMultimodalityPoint Cloud
🎯 What it does: Propose a deep learning architecture named RAVEN based on frequency-modulated continuous-wave radar, which processes radar ADC data in a streaming manner while preserving the spatial structure of the MIMO array, enabling subframe-level early decision-making for radar frames and outputting BEV space free-space segmentation and object detection;
RAW-Domain Degradation Models for Realistic Smartphone Super-Resolution
Ali Mosleh (Samsung Electronics), Alex Levinshtein (Samsung Electronics)
Super ResolutionConvolutional Neural NetworkImage
🎯 What it does: This paper addresses the super-resolution problem in the RAW domain of mobile phone cameras by constructing device-specific degradation models. These models are used in a simulation pipeline to generate synthetic RAW-LR/HR image pairs, which train a single-image super-resolution (SR) network from RAW to sRGB.
RawMetaDiff: Unlocking Extreme Darkness from Dual-Exposure RAW with Meta-Guided Diffusion
Panjun Liu (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
RestorationDiffusion modelAuto EncoderImage
🎯 What it does: Propose a conditional diffusion framework named RawMetaDiff, which recovers high-quality Raw images in extremely dark scenes using short-exposure Raw, long-exposure Raw with potential mismatch, and Raw metadata.
RAYNOVA: Scale-Temporal Autoregressive World Modeling in Ray Space
Yichen Xie (Applied Intuition), Wei Zhan (Applied Intuition)
GenerationAutonomous DrivingTransformerAuto EncoderWorld ModelImageVideoTextMultimodality
🎯 What it does: Designed and implemented a ray-space-based dual-causal autoregressive world model called RAYNOVA, capable of generating physically plausible videos in multi-perspective, multi-scale, and multi-frame driving scenarios. It supports various control signals, including text, object bounding boxes, and maps, while enabling image synthesis from novel viewpoints.
RC-NF: Robot-Conditioned Normalizing Flow for Real-Time Anomaly Detection in Robotic Manipulation
Shijie Zhou (Fudan University), Yu-Gang Jiang (Fudan University)
Anomaly DetectionRobotic IntelligenceTransformerFlow-based ModelPoint CloudBenchmark
🎯 What it does: Proposed a real-time robotic anomaly detection model, RC-NF, for monitoring whether the robot's execution status and target object trajectory conform to the task.
RDF-MIG: A Robust Diffusion Framework for Masked Image Generation to Augment Semantic Segmentation and Change Detection
Zian Cao (Huazhong University of Science and Technology), Yuanyuan Fu (Ping An Property & Casualty Insurance Company of China Ltd)
SegmentationGenerationData SynthesisConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Proposed RDF-MIG, a joint diffusion framework that simultaneously generates segmentation and change annotations, and improves generation quality through the MCRD loss.
RDFace: A Benchmark Dataset for Rare Disease Facial Image Analysis under Extreme Data Scarcity and Phenotype-Aware Synthetic Generation
Ganlin Feng (Western University), Pingzhao Hu (Western University)
ClassificationGenerationData SynthesisConvolutional Neural NetworkTransformerVision Language ModelDiffusion modelGenerative Adversarial NetworkImageBiomedical DataBenchmark
🎯 What it does: Constructed the RDFace benchmark dataset of facial images for rare diseases, and conducted experiments based on supervised learning, few-shot learning, and synthetic data augmentation using generative models, evaluating the diagnostic performance of different models under extremely low sample conditions.
Re-Align: Structured Reasoning-guided Alignment for In-Context Image Generation and Editing
Runze He (Tencent), Jiao Dai (Chinese Academy Of Sciences)
GenerationReinforcement LearningPrompt EngineeringVision Language ModelRectified FlowImageTextMultimodalityChain-of-Thought
🎯 What it does: Propose a unified framework Re-Align, leveraging structured reasoning (In-Context Chain-of-Thought, IC-CoT) to drive image generation and editing, achieving precise understanding and execution of complex interactive image-text prompts.
Re-evaluating Continual VQA: Toward Fair and Robust Evaluation for Multimodal Continual Learning
Zijian Gao (National University of Defense Technology), Huaimin Wang (National University of Defense Technology)
Knowledge DistillationTransformerVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposes a fairer and more robust multi-modal continual visual question answering (Continual VQA) evaluation framework, UCo-VQA, and designs a lightweight continual learning method, MaDQ, which enhances knowledge retention and visual-semantic alignment through question replay and dual-layer distillation;
RE-VLM: Event-Augmented Vision-Language Model for Scene Understanding
Hanqing Liu (Beijing University of Posts and Telecommunications), Chuang Zhu (Beijing University of Posts and Telecommunications)
RecognitionConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: Proposed RE-VLM, a dual-stream vision-language model that integrates RGB images and event streams, enabling robust scene understanding under adverse conditions such as poor illumination or high-speed motion.
REACH: Explicit Recovery Behavior for Diffusion Policies
Zundong Ke (ShanghaiTech University), Jiayuan Gu (ShanghaiTech University)
Robotic IntelligencePrompt EngineeringDiffusion modelAuto EncoderMultimodality
🎯 What it does: Propose the REACH framework, integrating error detection and negative prompting into diffusion policies to enable self-correction in robots during execution, enhancing robustness.
Reading or Reasoning? Format Decoupled Reinforcement Learning for Document OCR
Yufeng Zhong (Meituan), Lin Ma (Meituan)
RecognitionSupervised Fine-TuningReinforcement LearningMultimodalityBenchmark
🎯 What it does: Proposed and implemented a document OCR framework based on format-decoupled reinforcement learning (FD-RL), achieving high-quality end-to-end recognition of text, formulas, and tables through two-stage SFT+RL training.
Reading Your Actions: Learning Generalizable Action Representations via Pre-training AEMG
Zhenghao Huang (South China University of Technology), Lin Shu (South China University of Technology)
Representation LearningTransformerBiomedical Data
🎯 What it does: Construct a unified EMG representation model AEMG through large-scale self-supervised pre-training, converting EMG signals into 'neural words' and learning cross-device, cross-task general representations.
ReAG: Reasoning-Augmented Generation for Knowledge-based Visual Question Answering
Alberto Compagnoni (University of Modena and Reggio Emilia), Rita Cucchiara (University of Modena and Reggio Emilia)
GenerationRetrievalLarge Language ModelReinforcement LearningVision Language ModelMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a multi-modal retrieval-augmented generation framework ReAG for knowledge-driven visual question answering (KB-VQA) tasks;
ReaGEN: Adaptive Generation of Structured Chains-of-Thought for Efficient Multimodal Reasoning
Ruiqing Tian (Huawei Technologies), Bahador Rashidi (Huawei Technologies)
Explainability and InterpretabilityComputational EfficiencyTransformerVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Proposes the ReaGEN framework, which acquires sample-specific structured chain-of-thought (CoT) through teacher-guided evolutionary search, and trains a lightweight generator (GEN) to generate adaptive reasoning chains without altering the underlying vision-language model, thereby enhancing multimodal reasoning performance.
Real-Time Dynamic Scene Rendering with Controlled Compressibility and Contact Awareness
Boya Shi (Shanghai Jiao Tong University), Xiaodong Yi (Academy of Military Science)
GenerationComputational EfficiencyGaussian SplattingVideo
🎯 What it does: Propose a real-time dynamic scene rendering framework that integrates compressibility, source term modeling, and contact constraints, achieving physically consistent and temporally coherent rendering through velocity field projection of Gaussian primitives.
Real-Time Generation of Streamable Talking Portrait Video with Reference-Guided Deep Compression VAEs
Sicheng Xu (Microsoft Research), Baining Guo (Microsoft Research)
GenerationData SynthesisCompressionTransformerRectified FlowAuto EncoderImageVideoMultimodalityAudio
🎯 What it does: Developed a complete real-time streaming portrait video generation framework capable of synthesizing high-quality speaking portrait videos with a resolution of 512×512 and a frame rate of 42 fps based on speech audio and one or more reference images.
Real-Time Long Horizon Air Quality Forecasting via Group-Relative Policy Optimization
Inha Kang (KAIST AI), Hyunjung Shim (KAIST AI)
Convolutional Neural NetworkSupervised Fine-TuningReinforcement LearningTabularTime Series
🎯 What it does: Proposed a real-time long-term air quality prediction framework, FAKER-Air, which integrates CMAQ-OBS regional data and a two-stage training approach (SFT+GRPO) to achieve high-accuracy PM2.5/PM10 prediction.
Real-Time Multimodal Fingertip Contact Detection via Depth and Motion Fusion for Vision-Based Human-Computer Interaction
Mukhiddin Toshpulatov (KAIST), Geehyuk Lee (KAIST)
Object DetectionDepth EstimationTransformerSupervised Fine-TuningMultimodality
🎯 What it does: Utilizes a single camera combined with depth estimation and motion fusion to achieve real-time finger contact detection for VR keyboard input.
Real-Time Neural Video Compression with Unified Intra and Inter Coding
Hui Xiang (University of Science and Technology of China), Dong Liu (University of Science and Technology of China)
CompressionAuto EncoderVideo
🎯 What it does: Propose a unified neural video compression framework that integrates adaptive intra and inter encoding, and achieves real-time efficient compression through dual-frame synchronous compression.
Real-World Point Tracking with Verifier-Guided Pseudo-Labeling
Görkay Aydemir, Weidi Xie (Shanghai Jiao Tong University)
Object TrackingRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningVideo
🎯 What it does: Propose a Verifier meta-model that evaluates the reliability of candidate trajectories from multiple pre-trained point trackers frame by frame, generating high-quality pseudo labels for self-supervised fine-tuning on unlabeled videos.
Real2Edit2Real: Generating Robotic Demonstrations via a 3D Control Interface
Yujie Zhao (CFCS, School of Computer Science, Peking University), Hao Dong (CFCS, School of Computer Science, Peking University)
GenerationData SynthesisRobotic IntelligenceTransformerVision-Language-Action ModelImageVideoPoint Cloud
🎯 What it does: Deeply reliable spatial editing of multi-perspective robot demonstrations via a 3D editable interface, and synthesizing diverse, physically consistent demonstration videos using a 3D-controlled Transformer video generation model, significantly improving data collection efficiency.
Real2Sim2Real: RetinalDepth-64K for Depth Estimation in Posterior Segment Ophthalmic Surgery
Bingwen Dong (SUSTech), Jiang Liu (SUSTech)
Data SynthesisDepth EstimationSupervised Fine-TuningVideoTime SeriesBiomedical Data
🎯 What it does: Propose the RetinalDepth dataset and design the Real2Sim2Real pipeline for depth estimation in posterior segment ophthalmic surgery.
RealAppiance: Let High-fidelity Appliance Assets Controllable and Workable as Aligned Real Manauls
Yuzheng Gao, Hao Dong
Large Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Constructed 100 high-fidelity home appliance digital assets aligned with real user manuals, and proposed the RealAppliance-Bench evaluation benchmark based on this.
RealBirdID: Benchmarking Bird Species Identification in the Era of MLLMs
Logan Lawrence (UMass Amherst), Grant Van Horn (UMass Amherst)
ClassificationRecognitionTransformerLarge Language ModelVision Language ModelImageBenchmark
🎯 What it does: Proposed the RealBirdID benchmark to evaluate the refusal prediction and rationale generation capabilities of multimodal large models in fine-grained bird identification.
ReAlign: Generalizable Image Forgery Detection via Reasoning-Aligned Representation
Qing Huang (Peking University), Jian Zhang (Peking University)
Anomaly DetectionLarge Language ModelReinforcement LearningVision Language ModelContrastive LearningImageText
🎯 What it does: By performing contrastive learning on the reasoning text generated by large language models, transferring their knowledge to a lightweight CLIP detector, achieving generalizable detection of AI-generated images.
Realiz3D: 3D Generation Made Photorealistic via Domain-Aware Learning
Ido Sobol (Technion), Or Litany (Technion)
GenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelImageMesh
🎯 What it does: Lightweight fine-tuning of diffusion models to maintain realism with real images while learning 3D control from synthetic data.
Reallocating Attention Across Layers to Reduce Multimodal Hallucination
Haolang Lu (Beijing University of Posts and Telecommunications), Kun Wang (Nanyang Technological University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerMultimodalityBenchmark
🎯 What it does: Propose a lightweight, training-agnostic plugin that first identifies shallow perceptual heads and deep reasoning heads via a functional head, then applies category-conditioned rescaling to them, thereby mitigating hallucination issues in multi-modal large reasoning models without altering the model architecture.
REALM: An MLLM-Agent Framework for Open World 3D Reasoning Segmentation and Editing on Gaussian Splatting
Changyue Shi (Hangzhou Dianzi University), Zhou Yu (Hangzhou Dianzi University)
SegmentationTransformerLarge Language ModelVision Language ModelGaussian SplattingMultimodalityBenchmark
🎯 What it does: Proposed the REALM framework, utilizing multimodal large language models (MLLM) to achieve open-vocabulary 3D reasoning segmentation and editing in 3D Gaussian Splatting.
RealUnify: Do Unified Models Truly Benefit from Unification? A Comprehensive Benchmark
Yang Shi (Peking University), Ziwei Liu (Nanyang Technological University)
GenerationMultimodalityBenchmark
🎯 What it does: This paper proposes the RealUnify benchmark for systematically evaluating the bidirectional synergy between two core capabilities of unified models: visual understanding and generation;
RealVLG-R1: A Large-Scale Real-World Visual-Language Grounding Benchmark for Robotic Perception and Manipulation
Linfei Li (Tongji University), Ying Shen (Tongji University)
Object DetectionSegmentationPose EstimationRobotic IntelligenceTransformerReinforcement LearningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed the RealVLG framework, combining the RealVLG-11B large-scale multi-granularity dataset with the RealVLG-R1 model to achieve unified zero-shot reasoning from natural language to target localization, segmentation, grasping pose, and contact points.
REArtGS++: Generalizable Articulation Reconstruction with Temporal Geometry Constraint via Planar Gaussian Splatting
Di Wu (Hefei Institutes Of Physical Science Chinese Academy Of Sciences), Cewu Lu (Shanghai Jiao Tong University)
Pose EstimationDepth EstimationGaussian SplattingImageMesh
🎯 What it does: Using multi-view RGB two-state images, combining component-oriented planar Gaussian splatting and unsupervised helical motion modeling to recover component-level meshes and joint parameters of unseen articulated objects.