CVPR 2026 Papers — Page 14
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
FreqEdit: Preserving High-Frequency Features for Robust Multi-Turn Image Editing
Yucheng Liao (Sun Yat-sen University), Xudong Mao (Sun Yat-sen University)
GenerationDiffusion modelRectified FlowImage
🎯 What it does: Proposes an untrained framework called FreqEdit to address quality degradation caused by the loss of high-frequency information in multi-round instruction-based image editing;
FreqSIC: Frequency-aware Stereo Image Compression with Bi-directional Checkerboard Context Model
Shiyu Qin (Tsinghua University), Bin Chen (Harbin Institute of Technology)
CompressionImage
🎯 What it does: This paper proposes FreqSIC, addressing the issues of high-frequency information loss and encoding delay caused by autoregressive models in stereo image compression. It designs a frequency domain context transfer module (FSCT) and a bidirectional checkerboard context model to achieve more efficient probabilistic modeling and faster encoding speed.
Frequency Switching Mechanism for Parameter-Efficient Multi-Task Learning
Shih-Wen Liu (National Cheng Kung University), Yu-Chiang Frank Wang (NVIDIA Research)
SegmentationComputational EfficiencyImageBenchmark
🎯 What it does: Proposes Free Sinewich, an efficient multi-task learning framework that achieves parameter sharing through a frequency switching mechanism.
Frequency-Aware Affinity for Weakly Supervised Semantic Segmentation
Ziqian Yang (Xi'an Jiaotong-Liverpool University), Jimin Xiao (Xi'an Jiaotong-Liverpool University)
SegmentationTransformerVision Language ModelImage
🎯 What it does: Propose the Dual Frequency-Aware framework, leveraging low-frequency alignment and high-frequency correction for frequency-aware associations, generating precise and complete CAMs to achieve single-stage training for weakly supervised semantic segmentation.
Frequency-Aware Flow Matching for High-Quality Image Generation
Sucheng Ren (Johns Hopkins University), Liang-Chieh Chen (ByteDance)
GenerationData SynthesisConvolutional Neural NetworkTransformerFlow-based ModelAuto EncoderImage
🎯 What it does: Proposed a frequency-aware flow matching model called FreqFlow, which significantly improves image generation quality by combining low-frequency and high-frequency branches with adaptive weights.
Frequency-domain Manipulation for Face Obfuscation
Jintae Kim (Korea University), Chang-Su Kim (Korea University)
Safty and PrivacyImage
🎯 What it does: Propose a facial stealth framework called FreM based on frequency domain subband adaptive modulation, which can erase human-visible identity while preserving machine-recognizable information;
Fresco: Frequency-Spatial Consistent Optimization for Fine-Grained Head Avatar Modeling
Shikun Zhang (Monash University), Cunjian Chen (Monash University)
GenerationOptimizationVideo
🎯 What it does: Proposed the Fresco optimization framework, combining frequency domain curriculum learning with UV space consistency constraints, significantly enhancing the detail quality and cross-view consistency of head animation models.
From 2D Alignment to 3D Plausibility: Unifying Heterogeneous 2D Priors and Penetration-Free Diffusion for Occlusion-Robust Two-Hand Reconstruction
Gaoge Han (AgiBot Mohamed bin Zayed University of Artificial Intelligence), Tongliang Liu (University of Sydney)
GenerationPose EstimationTransformerDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes a dual-hand 3D reconstruction framework based on a single image, which can simultaneously recover the pose and shape of both hands.
From 3D Pose to Prose: Biomechanics-Grounded Vision-Language Coaching
Yuyang Ji (Drexel University), Feng Liu (Drexel University)
Pose EstimationExplainability and InterpretabilityConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoBiomedical Data
🎯 What it does: Proposes BioCoach, a real-time fitness coaching framework that integrates 3D skeletal kinematics with visual language models to generate precise feedback based on biomechanics.
From Attraction to Equilibrium: Physics-Inspired Semantic Gravitons for Zero-Shot Anomaly Detection
Yuwen Pan (Tsinghua University), You He (Tsinghua University)
Anomaly DetectionTransformerVision Language ModelContrastive LearningImageTextBiomedical DataBenchmarkPhysics Related
🎯 What it does: Proposed a physics-based dynamics semantic graviton network (SGNet) for zero-shot anomaly detection and segmentation.
From Contrast to Consistency: Rethinking Event-based Continuous-Time Optical Flow Estimation
Rui Hu (Xidian University), Jinjian Wu (Xidian University)
Convolutional Neural NetworkOptical FlowTime SeriesBenchmark
🎯 What it does: Proposes a hybrid supervised framework for continuous-time optical flow estimation based on event cameras, and introduces spatiotemporal structural consistency constraints.
From Corners to Fiducial Tags: Revisiting Checkerboard Calibration for Event Cameras
Taehun Ryu (University of Science and Technology), Kyungdon Joo (University of Science and Technology)
OptimizationTime Series
🎯 What it does: This paper proposes a calibration framework directly based on event data, utilizing corner detection to achieve chessboard and AprilTag calibration without reconstruction, avoiding artifacts caused by traditional event-based image reconstruction.
From Detection to Association: Learning Discriminative Object Embeddings for Multi-Object Tracking
Yuqing Shao (East China University of Science and Technology), Xiao Sun (Shanghai AI Laboratory)
Object TrackingDepth EstimationTransformerContrastive LearningVideo
🎯 What it does: Propose FDTA, an end-to-end multi-object tracking framework based on DETR, which explicitly enhances the discriminativeness of object embeddings through three adapters (spatial, temporal, and identity), significantly improving association accuracy.
From Events to Clarity: The Event-Guided Diffusion Framework for Dehazing
Ling Wang (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
RestorationDiffusion modelAuto EncoderMultimodality
🎯 What it does: Leverage the high dynamic range information captured by event cameras to implement haze removal in foggy images using conditional diffusion models.
From Exploration to Exploitation: A Two-Stage Entropy RLVR Approach for Noise-Tolerant MLLM Training
Donglai Xu (Independent Researcher), Lai-Man Po (City University of Hong Kong)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: Propose a two-stage RLVR training method based on token entropy to enhance the robustness of multimodal large language models under noisy annotations.
From Failure to Feedback: Group Revision Unlocks Hard Cases in Object-Level Grounding
Yuyuan Liu (University of Oxford), Junde Wu (University of Oxford)
Object DetectionSegmentationTransformerReinforcement LearningVision Language ModelImageText
🎯 What it does: Propose a group revision-based training paradigm that generates revised response groups from failed answers and uses reward shaping to recover gradient signals for hard samples;
From Feature Learning to Spectral Basis Learning: A Unifying and Flexible Framework for Efficient and Robust Shape Matching
Feifan Luo (Zhejiang University), Hongyang Chen (Zhejiang Lab)
RecognitionGraph Neural NetworkPoint CloudMesh
🎯 What it does: This paper proposes a learnable spectral basis function framework called Advanced Functional Maps, and realizes the first unsupervised spectral basis learning method, which can end-to-end simultaneously optimize shape features and spectral bases, thus achieving robust non-rigid 3D shape matching.
From Few-way to Many-way: Rethinking Few-shot Fine-grained Image Classification
Li-Jun Zhao (Shandong University), Xin-Shun Xu (Shandong University)
ClassificationRecognitionMeta LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an SCEG method for few-shot fine-grained image classification, expanding the problem for the first time from traditional few-way scenarios to more challenging many-way scenarios.
From Indoor to Open World: Revealing the Spatial Reasoning Gap in MLLMs
Mingrui Wu (University of Chinese Academy of Sciences), Tong Zhang (University of Chinese Academy of Sciences)
Object DetectionObject TrackingVideoMultimodalityPoint CloudBenchmark
🎯 What it does: Proposes OSi-Bench, an outdoor spatial reasoning benchmark based on multi-sensor pedestrian perspective videos, evaluating relational, metric, and motion reasoning capabilities of multi-modal large language models (MLLM).
From Infusion to Assimilation Distillation for Medical Image Segmentation
Jiankang Hong (Tongji University), Junsong Yuan (State University of New York at Buffalo)
SegmentationKnowledge DistillationContrastive LearningImageBiomedical Data
🎯 What it does: Designed a two-phase knowledge distillation framework named IAD for compressing and enhancing medical image segmentation models.
From Inpainting to Layer Decomposition: Repurposing Generative Inpainting Models for Image Layer Decomposition
Jingxi Chen (University of Maryland), Yiannis Aloimonos (University of Maryland)
RestorationGenerationTransformerSupervised Fine-TuningDiffusion modelImageMultimodality
🎯 What it does: This paper proposes a lightweight image layer decomposition method that utilizes a pre-trained image restoration model to achieve foreground extraction and background recovery.
From Intuition to Investigation: A Tool-Augmented Reasoning MLLM Framework for Generalizable Face Anti-Spoofing
Haoyuan Zhang (University of Chinese Academy of Sciences), Zhen Lei (University of Chinese Academy of Sciences)
Anomaly DetectionTransformerReinforcement LearningVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: This paper proposes the Tool-Augmented Reasoning FAS (TAR-FAS) framework, which leverages multimodal large language models (MLLM) to perform fine-grained visual reasoning for face anti-spoofing (FAS) by integrating Chain-of-Thought with visual tools;
From Manuals to Actions: A Unified VLA Model for Chain-of-Thought Manual Generation and Robotic Manipulation
Chenyang Gu (Peking University), Shanghang Zhang (Peking University)
Robotic IntelligenceVision-Language-Action ModelDiffusion modelGaussian SplattingMultimodalityChain-of-Thought
🎯 What it does: Propose a unified Vision-Language-Action (VLA) model called ManualVLA, which first generates a multimodal operation manual and then converts it into robot executable actions;
From Measurement to Mitigation: Quantifying and Reducing Identity Leakage in Image Representation Encoders with Linear Subspace Removal
Daniel George (Persona Identities), Yifei Zhang (Persona Identities)
Safty and PrivacyRepresentation LearningTransformerDiffusion modelContrastive LearningImageBenchmark
🎯 What it does: Evaluate identity leakage in frozen visual encoders (e.g., CLIP, DINOv2/3, SSCD) from an attacker's perspective and propose a one-time linear subspace elimination method (ISP) to mitigate identity leakage.
From None to All: Self-Supervised 3D Reconstruction via Novel View Synthesis
Ranran Huang (Imperial College London), Krystian Mikolajczyk (Imperial College London)
GenerationPose EstimationDepth EstimationTransformerGaussian SplattingImage
🎯 What it does: This paper proposes NAS3R, a fully self-supervised forward framework that can jointly predict 3D Gaussian primitives and camera intrinsic and extrinsic parameters without any 3D or camera annotations, and without relying on pre-trained priors, enabling 3D geometry reconstruction and novel view synthesis from uncalibrated, unoriented multi-view images.
From Observation to Action: Latent Action-based Primitive Segmentation for VLA Pre-training in Industrial Settings
Jiajie Zhang (ShanghaiTech University), Alexander Kleiner (Hangzhou Dianzi University)
SegmentationTransformerVision-Language-Action ModelVideo
🎯 What it does: Proposed a fully unsupervised pipeline capable of automatically extracting and segmenting semantically coherent action primitives from continuous industrial videos, outputting segmented videos and corresponding latent action sequences directly usable for Vision-Language-Action pretraining.
From Pairs to Sequences: Track-Aware Policy Gradients for Keypoint Detection
Yepeng Liu (Wuhan University), Yongchao Xu (Xiaomi EV)
Object TrackingConvolutional Neural NetworkReinforcement LearningSequential
🎯 What it does: Propose a trajectory-aware keypoint detection framework TraqPoint based on reinforcement learning, directly optimizing the long-term trackability of keypoints on image sequences;
From Panel to Pixel: Zoom-In Vision-Language Pretraining from Biomedical Scientific Literature
Kun Yuan (University of Strasbourg, CNRS, INSERM, ICube, UMR7357, Strasbourg, France), Serena Yeung-Levy (Stanford University)
ClassificationRetrievalRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningMultimodalityBiomedical Data
🎯 What it does: Design and implement the Panel2Patch pipeline to automatically parse multi-panel biomedical images and their textual descriptions into three-tiered vision-language pairs at the figure, panel, and patch levels, and perform hierarchical pre-training based on this.
From Pixel to Precision: Enhancing Handwritten Mathematical Expression Recognition with Image-Level Reward
Ze Liu (University of Science and Technology of China), Qi Liu (University of Science and Technology of China)
RecognitionTransformerReinforcement LearningImage
🎯 What it does: Propose a reward mechanism based on Image Matching Score (IMS) and construct the IMPO framework, which directly optimizes visual consistency in the HMER model through reinforcement learning (GRPO).
From Rays to Projections: Better Inputs for Feed-Forward View Synthesis
Zirui Wu (Hong Kong University of Science and Technology), Jie Song (Hong Kong University of Science and Technology)
GenerationDepth EstimationTransformerAuto EncoderImagePoint Cloud
🎯 What it does: Proposes projective conditioning, replacing traditional Plücker ray encoding with target viewpoint point cloud projection images to achieve feedforward novel view synthesis;
From Scale to Speed: Adaptive Test-Time Scaling for Image Editing
Xiangyan Qu (Institute of Information Engineering, Chinese Academy of Sciences), Yujun Cai (University of Queensland)
Image TranslationComputational EfficiencyLarge Language ModelVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: Proposed ADE-CoT, an image editing method based on adaptive testing-time expansion, which significantly accelerates inference while maintaining or improving editing quality.
From Selection to Scheduling: Federated Geometry-Aware Correction Makes Exemplar Replay Work Better under Continual Dynamic Heterogeneity
Zhuang Qi, Xiangxu Meng (Harbin Institute Of Technology)
ClassificationFederated LearningKnowledge DistillationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: In federated continual learning, the example replay method is improved by combining geometric structure alignment (GSA) and energy-based geometric correction (EGC) to mitigate catastrophic forgetting caused by client heterogeneity and task-level imbalance.
From Sketch to Fresco: Efficient Diffusion Transformer with Progressive Resolution
Shikang Zheng (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
GenerationTransformerDiffusion modelImageVideoText
🎯 What it does: Proposes a training-free progressive resolution sampling framework called Fresco, which achieves high-resolution generation by utilizing a unified noise field and variance-guided progressive upsampling.
From Softmax to Dirichlet: Evidential Learning for Semi-supervised Semantic Segmentation
Huayu Mai (University Of Science And Technology Of China), Yuan Wang (University Of Science And Technology Of China)
SegmentationImage
🎯 What it does: Propose a semi-supervised semantic segmentation framework based on evidence learning, which models class probabilities using Dirichlet distribution and introduces Hyper-Evidential (HESS) to achieve more reliable pseudo-label selection, thereby enhancing model generalization ability.
From Spots to Pixels: Dense Spatial Gene Expression Prediction from Histology Images
Ruikun Zhang (Beijing Institute of Technology), Liyuan Pan (Beijing Institute of Technology)
GenerationConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: Propose PixNet, a dense prediction-based network that directly generates high-resolution gene expression maps from histopathological slice images, and predicts gene expression of spots of arbitrary size and scale by aggregating expression values within specified circular regions.
From Static to Dynamic: Exploring Self-supervised Image-to-Video Representation Transfer Learning
Yang Liu (University of Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
Domain AdaptationRepresentation LearningTransformerImageVideo
🎯 What it does: This paper proposes the Co-Settle framework, which performs self-supervised image-to-video representation transfer learning using a lightweight projection layer on a frozen image pre-trained encoder;
From Weights to Concepts: Data-Free Interpretability of CLIP via Singular Vector Decomposition
Francesco Gentile (University of Trento), Elisa Ricci (University of Trento)
Explainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a completely data-free and training-free interpretation framework called SITH, which directly analyzes attention heads in the weight space of CLIP's vision transformer.
From Where Things Are to What They Are For: Benchmarking Spatial-Functional Intelligence in Multimodal LLMs
Le Zhang (Mila - Quebec AI Institute, Universite De Montreal), Bo-Hsiang Tseng
TransformerLarge Language ModelVideoBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes SFI-Bench, a video-based benchmark for evaluating the capabilities of multi-modal large language models in two cognitive levels: structured spatial reasoning and functional reasoning.
FSFSplatter: Geometrically Accurate Reconstruction with Free Sparse-view Images within 2 minutes
Yibin Zhao, Jianjun Yi
Depth EstimationOptimizationComputational EfficiencyTransformerGaussian SplattingImage
🎯 What it does: Propose the FSFSplatter framework to achieve fast, geometrically accurate free sparse view reconstruction.
FSLoRA: Harmonizing Detection and Re-Identification via Freq-Spatial Low-Rank Adapter for One-Stage Person Search
Yanling Tian (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)
RecognitionObject DetectionRetrievalTransformerMixture of ExpertsImageVideoBenchmark
🎯 What it does: Designed and implemented FSLoRA (spectral-spatial low-rank adapter), which inserts adaptive spatial and frequency modules into the backbone network of a single-stage person re-identification framework to achieve layer-wise separation of task features, resolving conflicts between detection and re-identification subtasks.
Fuel Gauge: Estimating Chain-of-Thought Length Ahead of Time in Large Multimodal Models
Yuedong Yang (University of Texas at Austin), Radu Marculescu (University of Texas at Austin)
Computational EfficiencyLarge Language ModelMultimodalityChain-of-Thought
🎯 What it does: Propose the Fuel Gauge method to predict the chain-of-thought (CoT) length in large-scale multimodal models and control the generation process in advance.
Fully Decentralized Certified Unlearning
Hithem Lamri (New York University Abu Dhabi), Michail Maniatakos (New York University Abu Dhabi)
Federated LearningImage
🎯 What it does: Studied the algorithm RR-DU for achieving certified machine forgetting in fully decentralized networks with fixed topologies, providing theoretical guarantees and experimental validation.
FUN REC * Reconstructing Functional 3D Scenes from Egocentric Interaction Videos
Alexandros Delitzas, Daniel Barath
Object DetectionObject TrackingSegmentationGenerationPose EstimationDepth EstimationOptimizationVision Language ModelVideoPoint CloudMesh
🎯 What it does: Reconstructing interactive 3D digital twin scenes from single-shot first-person RGB-D interactive videos, automatically discovering and tracking joint movements to generate articulated 3D objects directly manipulable in simulations.
Functional Mean Flow in Hilbert Space
Zhiqi Li (Georgia Institute of Technology), Bo Zhu (Georgia Institute of Technology)
GenerationData SynthesisFlow-based ModelImageMeshTime SeriesStochastic Differential Equation
🎯 What it does: Proposes Functional Mean Flow (FMF), a one-step flow matching generative model implemented in an infinite-dimensional Hilbert space, applicable to continuous function generation tasks such as time series, images, PDEs, and 3D shapes.
FunFact: Building Probabilistic Functional 3D Scene Graphs via Factor-Graph Reasoning
Zhengyu Fu (ETH Zuerich), Zuria Bauer (ETH Zuerich)
Data SynthesisDepth EstimationOptimizationRepresentation LearningLarge Language ModelVision Language ModelImageGraph
🎯 What it does: Construct an open-vocabulary functional 3D scene graph based on pose RGB-D images, and obtain calibrated functional relationship confidence through dual-factor graph joint inference.
FUSAR-GPT: A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery
Xiaokun Zhang (Fudan University), Haipeng Wang (Fudan University)
ClassificationObject DetectionTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: Constructed the first SAR image-text-AlphaEarth feature triplet dataset and proposed FUSAR-GPT, a vision-language model specifically designed for SAR images, leveraging geospatial priors and spatiotemporal anchors to achieve feature compensation, and employing a two-stage decoupled SFT to separate knowledge injection from task execution.
FUSER: Feed-Forward Multiview 3D Registration Transformer and SE(3)$^N$ Diffusion Refinement
Haobo Jiang (Nanyang Technological University), Jianmin Zheng (Nanyang Technological University)
Pose EstimationComputational EfficiencyTransformerDiffusion modelImagePoint Cloud
🎯 What it does: The study proposes FUSER, a fully feed-forward multi-view 3D registration Transformer that directly predicts global pose in a unified latent space, avoiding traditional pairwise matching and synchronization steps.
Fusion in Your Way: Aligning Image Fusion with Heterogeneous Demands via Direct Preference Optimization
Weijian Su (Dalian University of Technology), Qiang Zhang (Dalian University of Technology)
Image HarmonizationGenerationVision Language ModelDiffusion modelAuto EncoderImageMultimodality
🎯 What it does: Proposes a DPO-based image fusion framework called DPOFusion that can adaptively generate fused images based on the heterogeneous needs of human and machine vision.
Fusion of Depth and Semantics for Probabilistic Floorplan Localization
Kecheng Ye (Chinese Academy of Sciences), Xu Yang (Chinese Academy of Sciences)
Pose EstimationDepth EstimationTransformerImage
🎯 What it does: Proposes a probabilistic ray framework that integrates depth and semantic information for indoor camera localization based on 2D floor plans.
FusionAgent: A Multimodal Agent with Dynamic Model Selection for Human Recognition
Jie Zhu (Michigan State University), Xiaoming Liu (Michigan State University)
RecognitionComputational EfficiencyTransformerReinforcement LearningAgentic AIVision Language ModelVideoMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the FusionAgent framework, which dynamically selects human recognition expert models using a multi-modal large language model and achieves adaptive score fusion through the Anchor-based Confidence Top-k (ACT) fusion method.
FusionRegister: Every Infrared and Visible Image Fusion Deserves Registration
Congcong Bian (Jiangnan University), Xiao-jun Wu (Jiangnan University)
TransformerOptical FlowImageMultimodality
🎯 What it does: Proposes a post-registration framework called FusionRegister based on visual priors to correct residual spatial mismatches during infrared-visible image fusion, thereby enhancing the structural accuracy and detail preservation of the fused results.
FVAR: Next-Focus Prediction for Visual Autoregressive Modeling
Xiaofan Li (Baidu Inc), Dingkang Liang (Baidu Inc)
GenerationKnowledge DistillationTransformerImage
🎯 What it does: Proposed FVAR, a visual autoregressive model based on next-focus prediction, replacing traditional scale downsampling, significantly reducing aliasing while enhancing details and text readability.
FVBench: Benchmarking Deepfake Video Detection Capability of Large Multimodal Models
Jiarui Wang (Shanghai Jiao Tong University), Xiongkuo Min (Shanghai Jiao Tong University)
ClassificationAnomaly DetectionConvolutional Neural NetworkTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: This paper constructs FVBench, a deepfake detection benchmark comprising 120k videos, and evaluates the performance of traditional detection models and large multilingual models (LMM) on real, AI-edited, and AI-generated videos.
G-MIXER: Geodesic Mixup-based Implicit Semantic Expansion and Explicit Semantic Re-ranking for Zero-Shot Composed Image Retrieval
Jiyoung Lim (Sungkyunkwan University), Jee-Hyong Lee (Sungkyunkwan University)
RetrievalTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Proposed a training-agnostic zero-shot synthetic image retrieval method called G-MIXER, which enhances implicit semantics through geodesic spherical mixing between image and text features and improves retrieval diversity and accuracy by leveraging explicit semantics for re-ranking.
G$^2$VLM: Geometry Grounded Vision Language Model with Unified 3D Reconstruction and Spatial Reasoning
Wenbo Hu (Shanghai AI Lab), Jiangmiao Pang (Shanghai AI Lab)
Pose EstimationDepth EstimationRepresentation LearningTransformerLarge Language ModelMixture of ExpertsVision Language ModelImagePoint Cloud
🎯 What it does: Propose the G VLM2 model, integrating 3D reconstruction and spatial understanding into a unified vision-language model.
GA-VLN: Geometry-Aware BEV Representation for Efficient Vision-Language Navigation
Jiahao Yang (Chinese Academy of Sciences), Shuqiang Jiang (University of Chinese Academy of Sciences)
Autonomous DrivingComputational EfficiencyTransformerLarge Language ModelVision Language ModelVision-Language-Action ModelImageMultimodalityPoint Cloud
🎯 What it does: Propose a Geometry-Aware BEV representation and integrate it into the visual language navigation framework GA-VLN, achieving efficient spatial perception and navigation;
Gallant: Voxel Grid-based Humanoid Locomotion and Local-navigation across 3-D Constrained Terrains
Qingwei Ben (Chinese University of Hong Kong), Jiangmiao Pang (Shanghai Artificial Intelligence Laboratory)
Domain AdaptationRobotic IntelligenceConvolutional Neural NetworkReinforcement LearningSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposes a voxel-grid-based perception framework called Gallant, achieving a unified perception and motion control strategy on three-dimensional constrained terrains.
Gamba: Mamba-based graph convolutional network with dynamic graph topology learning for action recognition
Rouyi Zhou (Shenzhen University), Linlin Shen (Shenzhen University)
RecognitionGraph Neural NetworkVideoBenchmark
🎯 What it does: This paper proposes the Gamba network, which combines Mamba and GCN to achieve dynamic graph topology learning and long-term temporal modeling for human action recognition.
GardenDesigner: Encoding Aesthetic Principles into Jiangnan Garden Construction via a Chain of Agents
Mengtian Li (Shanghai University), Zeyu Wang (Hong Kong University of Science and Technology (Guangzhou))
GenerationData SynthesisOptimizationLarge Language ModelAgentic AITextMeshRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the GardenDesigner framework, which automatically generates 3D layouts of Jiangnan gardens through text input and supports real-time editing and interaction in Unity.
Garments2Look: A Multi-Reference Dataset for High-Fidelity Outfit-Level Virtual Try-On with Clothing and Accessories
Junyao Hu (Hong Kong Polytechnic University), Xingxing Zou (Hong Kong Polytechnic University)
Image TranslationData SynthesisPose EstimationRetrievalTransformerLarge Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: This paper constructs the Garments2Look dataset, achieving a multi-reference virtual try-on scenario for complete clothing coordination.
Gastric-X: A Multimodal Multi-Phase Benchmark Dataset for Advancing Vision-Language Models in Gastric Cancer Analysis
Yuanzhe Li, Sheng Lu
ClassificationObject DetectionRetrievalTransformerSupervised Fine-TuningVision Language ModelImageMultimodalityTabularBiomedical DataComputed TomographyElectronic Health RecordsBenchmark
🎯 What it does: Propose the Gastric-X dataset, constructing a complete medical dataset containing 1.7K cases with multi-modal, multi-phase 3D CT, endoscopic images, structured biochemical indicators, diagnostic reports, and 3D lesion bounding boxes. Based on this dataset, five clinically relevant tasks (VQA, report generation, cross-modal retrieval, staging classification, lesion localization) are designed, providing a unified evaluation benchmark for multi-modal vision-language learning in gastric cancer diagnosis.
Gated Condition Injection without Multimodal Attention: Towards Controllable Linear-Attention Transformers
Yuhe Liu (National University Of Singapore), Xinchao Wang (National University Of Singapore)
GenerationTransformerSupervised Fine-TuningDiffusion modelImageTextMultimodality
🎯 What it does: Proposed a controllable diffusion framework based on a gating mechanism (GateControl) for the linear attention SANA model, enabling efficient control for subject-oriented generation and spatial alignment tasks.
Gated KalmaNet: A Fading Memory Layer through Test-time Ridge Regression
Liangzu Peng (University of Pennsylvania), Stefano Soatto (AWS Agentic AI)
ClassificationRetrievalImageText
🎯 What it does: Proposed Gated KalmaNet (GKA) — a decay memory layer based on Kalman filtering, which dynamically updates states during testing using online ridge regression, maintaining linear time and constant storage;
Gau-Occ: Geometry-Completed Gaussians for Multi-Modal 3D Occupancy Prediction
Chengxin Lv (Beihang University), YunHong Wang (Beihang University)
SegmentationAutonomous DrivingDiffusion modelGaussian SplattingImageMultimodalityPoint Cloud
🎯 What it does: Propose a multi-modal 3D semantic occupancy prediction framework that utilizes geometrically complete Gaussian primitives combined with sparse LiDAR to achieve functional and Gaussian anchor fusion modules, enabling efficient fusion of sparse LiDAR and multi-view images.
GauMVC: Generative Decoupled Gaussian Representation for Human-centric Multi-view Video Compression
Ruoke Yan (Peking University), Siwei Ma (Peking University)
CompressionGaussian SplattingVideo
🎯 What it does: This paper proposes a compression framework called GauMVC based on generative separation, which divides human-centered multi-view videos into static background and dynamic human components. The static background and dynamic human parts are encoded separately using 3D Gaussian fields, SMPL parameters, and sparse key views. The decoder achieves high-quality free-viewpoint reconstruction by fusing SMPL-driven Gaussian human models with the background.
GaussFusion: Improving 3D Reconstruction in the Wild with A Geometry-Informed Video Generator
Liyuan Zhu (Stanford University), Iro Armeni (Stanford University)
GenerationKnowledge DistillationConvolutional Neural NetworkTransformerDiffusion modelFlow-based ModelAuto EncoderGaussian SplattingImageVideo
🎯 What it does: Propose GaussFusion, a geometry information-driven video generation model that refines 3D Gaussian splatting (3DGS) reconstruction results, removing artifacts such as floating points, flickering, and blurriness, and further optimizing the underlying 3D representation.
Gaussian Mapping for Evolving Scenes
Vladimir Yugay (University of Amsterdam), Lukas Schmid (University of Amsterdam)
GenerationPose EstimationDepth EstimationGaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: Construct an online-updatable 3D Gaussian Splatting (3DGS) map, real-time detect and process long-term dynamic changes in the environment (adding, moving, removing geometries), achieving high-quality novel view synthesis;
Gaussian Splatting-based Low-Rank Tensor Representation for Multi-Dimensional Image Recovery
Yiming Zeng (University of Electronic Science and Technology of China), Chao Wang (Southern University of Science and Technology)
RestorationGaussian SplattingImage
🎯 What it does: Propose a low-rank tensor representation based on Gaussian splatting (GSLR) for multidimensional image recovery, generating the latent tensor with 2D Gaussian points and the transformation matrix with 1D Gaussian points, and building an unsupervised recovery model based on this.
Gaussian-Mixture Latent Flow for Stochastic 3D Human Motion Prediction
Yue Ma (Beihang University), Xiaohui Liang (Beihang University)
Pose EstimationTransformerFlow-based ModelTime SeriesSequential
🎯 What it does: Proposed a random 3D human motion prediction framework based on Gaussian Mixture Implicit Flows.
GaussianDWM: 3D Gaussian Driving World Model for Unified Scene Understanding and Multi-Modal Generation
Tianchen Deng (Shanghai Jiao Tong University), Hesheng Wang (Shanghai Jiao Tong University)
Autonomous DrivingLarge Language ModelVision Language ModelDiffusion modelAuto EncoderGaussian SplattingWorld ModelImageMultimodality
🎯 What it does: Developed a unified driving world model based on 3D Gaussian fields (GaussianDWM), achieving 3D perception and multimodal generation in driving scenarios.
GaussianFluent: Gaussian Simulation for Dynamic Scenes with Mixed Materials
Bei Huang (Peking University), Siyuan Huang (BIGAI)
GenerationData SynthesisGaussian SplattingImage
🎯 What it does: GaussianFluent proposes a unified framework to achieve photorealistic dynamic scene simulation on 3D Gaussian Splatting, particularly for real-time rendering of brittle fracture and mixed material objects.
GaussianGrow: Geometry-aware Gaussian Growing from 3D Point Clouds with Text Guidance
Weiqi Zhang (Tsinghua University), Yu-Shen Liu (Tsinghua University)
GenerationDiffusion modelGaussian SplattingPoint Cloud
🎯 What it does: Propose GaussianGrow, achieving text-guided high-quality 3D content generation by growing 3D Gaussians from point clouds.
GaussianMatch: Semi-Supervised Regression with Pseudo-Label Filtering via Multi-View Gaussian Consistency
Yin Wang (Zhejiang University), Shuiguang Deng (Zhejiang University)
OptimizationData-Centric LearningImageText
🎯 What it does: Proposed GaussianMatch, a framework for semi-supervised regression that utilizes multi-view prediction consistency to filter high-quality pseudo labels;
GaussianPile: A Unified Sparse Gaussian Splatting Framework for Slice-based Volumetric Reconstruction
Di Kong (Tsinghua University), Cheng Ma (Tsinghua University)
RestorationCompressionNeural Radiance FieldGaussian SplattingBiomedical DataUltrasound
🎯 What it does: Proposes the GaussianPile framework, combining 3D Gaussian points with slice imaging physics models to achieve efficient compression and reconstruction of slice-based volumetric data.
GaussianVision: Vision-Language Alignment from Compressed Image Representations using 2D Gaussian Splatting
Yasmine Omri, Thierry Tambe
CompressionRepresentation LearningLarge Language ModelVision Language ModelContrastive LearningGaussian SplattingImageTextMultimodality
🎯 What it does: This paper constructs an scalable 2DGS pipeline by using 2D Gaussian Splatting as the input of a vision-language model, and migrates the CLIP model to this representation, subsequently verifying its effectiveness in the VQA task.
GaussianZoom: Progressive Zoom-in Generative 3D Gaussian Splatting with Geometric and Semantic Guidance
Jiale Shi (Zhejiang University), Zhaopeng Cui (Zhejiang University)
GenerationSuper ResolutionDiffusion modelGaussian SplattingImage
🎯 What it does: Designed and implemented a progressive generative 3D Gaussian Splatting framework—GaussianZoom—for achieving extreme upscaling rendering from low-resolution images.
Gaze Target Estimation Anywhere with Concepts
Xu Cao (University Of Illinois Urbana Champaign), James M. Rehg (University Of Illinois Urbana Champaign)
Object DetectionTransformerPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Proposed an end-to-end gaze target estimation task (Promptable Gaze Target Estimation, PGE) that can be completed through natural language or visual prompts, and implemented the GazeAnywhere model.
GazeOnce360: Fisheye-Based 360deg Multi-Person Gaze Estimation with Global-Local Feature Fusion
Zhuojiang Cai (Beihang University), Feng Lu (Beihang University)
RecognitionData SynthesisConvolutional Neural NetworkImage
🎯 What it does: Proposed GazeOnce360, an end-to-end method for directly estimating 3D gaze directions of multiple people from a ceiling-mounted upward-facing fisheye camera, and constructed a large-scale synthetic dataset called MPSGaze360.
GazeShift: Unsupervised Gaze Estimation and Dataset for VR
Gil Shapira (Samsung Semiconductor Israel R&D Center), Niv Zehngut (Samsung Semiconductor Israel R&D Center)
Pose EstimationRepresentation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposed the VRGaze large-scale off-axis VR near-eye image dataset and designed the GazeShift unsupervised gaze estimation framework, which can learn high-quality gaze representations without annotated data;
GDFA: Geometry-Driven Federated Unlearning with Directional Task Vector Alignment
Xiuting Weng (Yunnan University), Lixing Yu (Yunnan University)
Federated LearningConvolutional Neural NetworkImage
🎯 What it does: Proposed a geometry-driven federated learning framework GDFA, achieving efficient deletion of specified client data in federated learning.
GDPO-SR: Group Direct Preference Optimization for One-Step Generative Image Super-Resolution
Qiaosi Yi (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
Super ResolutionReinforcement LearningDiffusion modelImage
🎯 What it does: Proposed a single-step generative image super-resolution framework GDPO (Group Direct Preference Optimization) based on reinforcement learning
GDRO: Group-level Reward Post-training Suitable for Diffusion Models
Yiyang Wang (University of Hong Kong), Hengshuang Zhao (University of Hong Kong)
GenerationOptimizationSupervised Fine-TuningReinforcement LearningDiffusion modelText
🎯 What it does: Propose GDRO, a post-training group-level reward optimization method for diffusion models, achieving offline optimization without sampler dependency using an implicit reward function.
GeCo-SRT: Geometry-aware Continual Adaptation for Cross-Task Sim-to-Real Transfer
Wenbo Yu (Beijing Forestry University), Di Hu (Renmin University of China)
Domain AdaptationRobotic IntelligenceMixture of ExpertsDiffusion modelPoint Cloud
🎯 What it does: Developed the GeCo-SRT method to achieve geometry-aware continuous cross-task sim-to-real transfer.
GeCo: Geometry-Consistent Regularization for Domain Generalized Semantic Segmentation
Qi Zang (Hefei University of Technology), Meng Wang (Hefei University of Technology)
SegmentationDomain AdaptationTransformerImage
🎯 What it does: This paper addresses the domain generalization semantic segmentation task by proposing Geometry-Consistent Regularization (GeCo), which significantly enhances the model's robustness and unknown category recognition capability on unseen domains through geometry-consistent regularization of lightweight adapters (e.g., LoRA) for Vision Foundation Models (VFMs).
GEM-TFL: Bridging Weak and Full Supervision for Forgery Localization through EM-Guided Decomposition and Temporal Refinement
Xiaodong Zhu (Wuhan University), Zhongyuan Wang (Wuhan University)
Object DetectionAnomaly DetectionTransformerVideo
🎯 What it does: This paper proposes a two-stage weakly supervised video forgery localization framework named GEM-TFL, which precisely locates forged segments using only video-level binary labels through EM-guided attribute decomposition and temporal consistency refinement.
GEM: Generating LiDAR World Model via Deformable Mamba
Yang Wu (NJUST), Jin Xie (Nanjing University)
Autonomous DrivingDiffusion modelGenerative Adversarial NetworkWorld ModelPoint Cloud
🎯 What it does: Proposes GEM—a generative LiDAR world model based on Mamba, which discretizes LiDAR scans using a customized LiDAR scene tokenizer, employs an unsupervised dynamic-static separator, and adaptively scans, aggregates, and fuses dynamic, static, and general features through a three-path deformable Mamba, achieving high-precision prediction and controllable generation of future LiDAR sequences;
Gen3R: 3D Scene Generation Meets Feed-Forward Reconstruction
Jiaxin Huang (Zhejiang University), Yiyi Liao (Zhejiang University)
GenerationData SynthesisPose EstimationDepth EstimationTransformerDiffusion modelAuto EncoderVideoPoint Cloud
🎯 What it does: Propose the Gen3R method, which integrates the VGGT reconstruction model with a video diffusion model to jointly generate RGB videos and corresponding 3D geometry (point clouds, depth maps, camera poses) while achieving multi-view controllability.
GenBreak: Red Teaming Text-to-Image Generation Using Large Language Models
Zilong Wang (Fudan University), Xingjun Ma (Fudan University)
GenerationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes GenBreak, a framework that utilizes large language models to perform red team attacks on text-image generation models.
GenColorBench: A Color Evaluation Benchmark for Text-to-Image Generation
Muhammad Atif Butt (Computer Vision Center), Kai Wang (City University of Hong Kong)
GenerationPrompt EngineeringVision Language ModelDiffusion modelImageTextBenchmark
🎯 What it does: Proposed the GenColorBench benchmark to evaluate the accuracy and controllability of text-to-image generation models in color generation.
General Process Reward Modeling for Robotic Reinforcement Learning
Huajie Tan (Peking University), Shanghang Zhang (Peking University)
Robotic IntelligenceTransformerReinforcement LearningVision Language ModelImageVideoTextMultimodality
🎯 What it does: This paper proposes a general reward model GRM (Dopamine-Reward) based on multi-perspective views and a theoretically optimal policy-preserving reward shaping framework Dopamine-RL, for efficiently learning long-horizon manipulation tasks in real robots.
Generalizable Co-Salient Object Detection via Mixed Content-Style Modulation
Guanting Guo (Nanjing University of Information Science and Technology), Min Xia (IEIT SYSTEMS Co., Ltd.)
Object DetectionDomain AdaptationConvolutional Neural NetworkTransformerVision Language ModelImageBenchmark
🎯 What it does: Proposed a CoMCS framework leveraging hybrid content-style modulation to enhance the generalization ability of co-salient object detection in unseen domains.
Generalizable Knowledge Distillation from Vision Foundation Models for Semantic Segmentation
Chonghua Lv (Xidian University), Zhun Zhong (Hefei University of Technology)
SegmentationKnowledge DistillationTransformerImageBenchmark
🎯 What it does: This paper studies how to distill the general knowledge of Vision Foundation Models (VFMs) into lightweight semantic segmentation models to enhance cross-domain generalization capabilities.
Generalizable Radio-Frequency Radiance Fields for Spatial Spectrum Synthesis
Kang Yang (University of California Merced), Wan Du (University of California Merced)
GenerationData SynthesisTransformerNeural Radiance FieldMeshPhysics Related
🎯 What it does: Proposed GRaF, a general RF radiation field framework that can generate spatial spectra at arbitrary positions in different scenarios based on the nearby spectra of known transmitters.
Generalizable Sparse-View 3D Reconstruction from Unconstrained Images
Vinayak Gupta (University of Maryland College Park), Jia-Bin Huang (University of Maryland College Park)
Image TranslationGenerationTransformerDiffusion modelGaussian SplattingImagePoint Cloud
🎯 What it does: Developed a method that generates controllable lighting and occlusion-adaptive 3D Gaussian structures from sparse, unposed outdoor internet images without scene optimization, achieving high-quality, view-consistent 3D reconstruction in about 3 seconds.
Generalizable Structure-Aware Keypoint Correspondence for Category-Unified 3D Single Object Tracking
Jie Xiao (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
Object TrackingPose EstimationAutonomous DrivingTransformerPoint Cloud
🎯 What it does: Proposes a unified 3D single-target tracking framework called UniKPT, which can simultaneously track multiple categories of targets within a single model.
Generalizable Video Quality Assessment via Weak-to-Strong Learning
Linhan Cao (Shanghai Jiao Tong University), Xiongkuo Min (Shanghai Jiao Tong University)
Domain AdaptationKnowledge DistillationData-Centric LearningVideoBenchmark
🎯 What it does: Explored an unsupervised weak-to-strong learning framework that trains stronger student models using pseudo-labels generated by weak VQA teachers, achieving video quality assessment without requiring human annotations.
Generalized and Personalized Federated Learning with Black-Box Foundation Models via Orthogonal Transformations
Eun Gyung Kong (Mobilint, Inc.), Taesup Kim (Seoul National University)
ClassificationFederated LearningSafty and PrivacyTransformerImage
🎯 What it does: This paper proposes FEDOT, a federated learning framework for black-box base models, which achieves a balance between personalization and generalization by using a globally shared classifier and local orthogonal transformations.
Generalized-CVO: Fast and Correspondence-Free Local Point Cloud Registration with Second Order Riemannian Optimization
Ray Zhang (Toyota Research Institute), John Subosits (Toyota Research Institute)
Pose EstimationAutonomous DrivingOptimizationPoint Cloud
🎯 What it does: Propose a correspondence-free point cloud registration method called G-CVO, which utilizes spatially varying anisotropic kernel embedding in RKHS and performs second-order Riemannian Gauss-Newton optimization on SE(3);
Generalizing Visual Geometry Priors to Sparse Gaussian Occupancy Prediction
Changqing Zhou (Hong Kong University of Science and Technology), Changhao Chen (Hong Kong University of Science and Technology)
SegmentationDepth EstimationGaussian SplattingImageVideo
🎯 What it does: This paper proposes a 3D occupancy prediction framework called GPOcc that utilizes visual geometric priors (GPs) for sparse Gaussian rendering.
GenErase: Generalizable and Semantically-Aware Concept Erasure in Diffusion Models
Korada Sri Vardhana (Indian Institute of Science), Soma Biswas (Indian Institute of Science)
GenerationDiffusion modelImageTextBenchmark
🎯 What it does: Proposes a training-free concept erasure framework called GenErase, which achieves precise erasure of target concepts in cross-attention value space through geometric constraints;