ICCV 2025 Papers — Page 11
IEEE/CVF International Conference on Computer Vision · 2701 papers
Generate, Transduct, Adapt: Iterative Transduction with VLMs
Oindrila Saha (University of Massachusetts), Subhransu Maji (University of Massachusetts)
ClassificationDomain AdaptationTransformerLarge Language ModelContrastive LearningImage
🎯 What it does: In the scenario of image classification with no labels or few labels, the GTA-CLIP framework is proposed, which utilizes large language models to dynamically generate category attributes, combining attribute-enhanced image-to-image transductive inference and adaptation of the CLIP encoder to form an iterative generation-transduction-adaptation process;
Generating Multi-Image Synthetic Data for Text-to-Image Customization
Nupur Kumari (Carnegie Mellon University), Samaneh Azadi (Meta)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes a dataset called SynCD, which generates synthesized images of the same object from multiple perspectives, lighting conditions, and backgrounds using 3D assets and text prompts. An encoder-based text-to-image custom model is trained on this dataset, ultimately achieving higher quality and better identity preservation in custom generation.
Generating Physically Stable and Buildable Brick Structures from Text
Ava Pun (Carnegie Mellon University), Jun-Yan Zhu (Carnegie Mellon University)
GenerationRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMesh
🎯 What it does: This paper proposes a framework called BRICKGPT based on autoregressive large language models, which can generate physically stable LEGO structures that can be assembled manually or by robots based on free text prompts, and supports color and texture extensions.
Generating, Fast and Slow: Scalable Parallel Video Generation with Video Interface Networks
Bhishma Dedhia (Princeton University), Yuchen Liu (Adobe Research)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelOptical FlowVideoBenchmark
🎯 What it does: Proposes a Video Interface Network (VIN) that achieves parallel generation of video segments through global semantic tagging, significantly reducing computational load and enhancing temporal consistency in long videos.
Generative Active Learning for Long-tail Trajectory Prediction via Controllable Diffusion Model
Daehee Park (Intelligent Systems and Learning Lab.), Kuk-Jin Yoon (Qualcomm Research)
GenerationAutonomous DrivingDiffusion modelTime Series
🎯 What it does: This paper proposes a generative active learning framework for trajectory prediction called GALTraj, which uses a controllable diffusion model for generative data augmentation of long-tail samples.
Generative Adversarial Diffusion
U-Chae Jun (Sookmyung Women's University), Jiwoo Kang (Sookmyung Women's University)
GenerationData SynthesisDiffusion modelGenerative Adversarial NetworkImageMultimodality
🎯 What it does: A unified generative framework called Generative Adversarial Diffusion (GAD) is proposed, which incorporates adversarial loss into the denoising process of latent diffusion models at each step, using a single U-Net to serve as both the generator and discriminator, enhancing training stability and image quality.
Generative Gaussian Splatting: Generating 3D Scenes with Video Diffusion Priors
Katja Schwarz (Meta Reality Labs), Peter Kontschieder (Meta Reality Labs)
GenerationData SynthesisDiffusion modelGaussian SplattingImageVideo
🎯 What it does: This study proposes Generative Gaussian Splatting (GGS), which combines a pre-trained implicit video diffusion model with a 3D Gaussian sphere representation to directly generate consistent 3D scenes from a small number of images with known poses.
Generative Modeling of Shape-Dependent Self-Contact Human Poses
Takehiko Ohkawa (Meta), Takaaki Shiratori (University of Tokyo)
GenerationPose EstimationTransformerDiffusion modelMesh
🎯 What it does: This paper proposes the shape-conditioned partial attention latent diffusion model PAPoseDiff for generating self-contact poses by constructing a large-scale precise human shape self-contact dataset Goliath-SC, and applies this prior to pose correction in single-view pose estimation.
Generative Video Bi-flow
Chen Liu (University College London), Tobias Ritschel (University College London)
GenerationData SynthesisAutonomous DrivingFlow-based ModelVideoOrdinary Differential Equation
🎯 What it does: A generative video model based on neural ODE flows is proposed—Video Bi-flow, which achieves streaming, unconditional video generation by learning the temporal evolution from the previous frame to the next frame and a denoising flow.
Generative Zoo
Tomasz Niewiadomski (Max Planck Institute for Intelligent Systems), Peter Kulits (Max Planck Institute for Intelligent Systems)
GenerationData SynthesisPose EstimationConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: A dataset of multi-species animal poses and shapes, GenZoo, consisting of millions of images, was generated using a conditional image generation model, and a 3D pose and shape regression model was trained with this dataset.
Generic Event Boundary Detection via Denoising Diffusion
Jaejun Hwang (Pohang University of Science and Technology), Minsu Cho (Pohang University of Science and Technology)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: A general event boundary detection method based on the denoising diffusion model (DiffGEBD) is proposed, which can generate diverse and reasonable boundary predictions from the same video.
GenFlow3D: Generative Scene Flow Estimation and Prediction on Point Cloud Sequences
Hanlin Li (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
GenerationAutonomous DrivingRecurrent Neural NetworkTransformerDiffusion modelPoint Cloud
🎯 What it does: Proposes GenFlow3D, which jointly estimates the scene flow and future scene flow of point cloud sequences, utilizing recurrent networks and diffusion models for end-to-end learning.
GenFlowRL: Shaping Rewards with Generative Object-Centric Flow in Visual Reinforcement Learning
Kelin Yu (University of Maryland), Ruohan Gao (University of Maryland)
Robotic IntelligenceReinforcement LearningDiffusion modelVideo
🎯 What it does: This work uses the generated object center flow as a reward prior, combined with sparse state rewards, to achieve closed-loop control in visual reinforcement learning, significantly enhancing the robustness and generalization of robots in fine-grained, contact-rich tasks.
GenHancer: Imperfect Generative Models are Secretly Strong Vision-Centric Enhancers
Shijie Ma (Tencent), Ying Shan (Tencent)
GenerationRepresentation LearningTransformerVision Language ModelRectified FlowImage
🎯 What it does: Two-stage fine-tuning of CLIP ViT is performed, utilizing a self-supervised reconstruction task from generative models to enhance its fine-grained visual representation.
GenHaze: Pioneering Controllable One-Step Realistic Haze Generation for Real-World Dehazing
Sixiang Chen (Hong Kong University of Science and Technology), Lei Zhu (Hong Kong University of Science and Technology)
RestorationGenerationLarge Language ModelSupervised Fine-TuningDiffusion modelImage
🎯 What it does: We propose GenHaze, a single-step controllable haze generation framework based on a pre-trained latent diffusion model, which can directly convert clear images into realistic hazy images and utilize the generated hazy images to perform lightweight fine-tuning on existing dehazing models, significantly enhancing real-world dehazing performance.
GenieBlue: Integrating both Linguistic and Multimodal Capabilities for Large Language Models on Mobile Devices
Xudong Lu (vivo AI Lab), Hongsheng Li (Chinese University of Hong Kong)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodality
🎯 What it does: This paper proposes GenieBlue, a large language model (LLM) structure that balances pure language and multimodal capabilities on mobile devices. By freezing the original LLM parameters, copying Transformer blocks every four layers, and adding LoRA to the remaining blocks, it achieves multimodal training without compromising text capabilities, and employs a non-shared benchmark deployment strategy. It is deployed and evaluated on the Qualcomm Snapdragon 8 Elite NPU.
GenM3: Generative Pretrained Multi-path Motion Model for Text Conditional Human Motion Generation
Junyu Shi (Hong Kong University of Science and Technology), Qiang Nie (Hong Kong University of Science and Technology)
GenerationData SynthesisPose EstimationTransformerMixture of ExpertsVideoTextMultimodality
🎯 What it does: This paper proposes the GenM3 framework, which utilizes a multi-expert VQ-VAE and multi-path Transformer to achieve a unified discrete representation of multi-source human actions.
GENMO: A GENeralist Model for Human MOtion
Jiefeng Li (NVIDIA), Ye Yuan (NVIDIA)
GenerationPose EstimationTransformerDiffusion modelVideoTextMultimodality
🎯 What it does: A general model GENMO is proposed, which unifies the estimation of human actions (recovering 3D actions from video/2D poses) and generation (generating actions under multimodal conditions such as music, text, video, keyframes, etc.), and supports variable-length sequences and multimodal mixed conditions.
Geo4D: Leveraging Video Generators for Geometric 4D Scene Reconstruction
Zeren Jiang (University of Oxford), Andrea Vedaldi (University of Oxford)
GenerationData SynthesisDepth EstimationOptimizationTransformerDiffusion modelAuto EncoderVideoPoint Cloud
🎯 What it does: Using a pre-trained video diffusion model, Geo4D predicts point clouds, disparity maps, and ray maps from monocular videos in a single pass, and then fuses them through a multi-modal alignment algorithm to obtain a complete 4D geometric reconstruction result.
GeoAvatar: Adaptive Geometrical Gaussian Splatting for 3D Head Avatar
SeungJun Moon (Klleon AI Research), Gyeong-Moon Park (Korea University)
GenerationData SynthesisGaussian SplattingVideo
🎯 What it does: Proposes the GeoAvatar framework, which uses adaptive geometric Gaussian projection to generate 3D head avatars that maintain identity consistency across multiple poses.
GEOBench-VLM: Benchmarking Vision-Language Models for Geospatial Tasks
Muhammad Danish, Salman Khan (Australian National University)
ClassificationObject DetectionSegmentationTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This paper presents GEOBench-VLM, a new benchmark specifically designed to evaluate visual-language models on geospatial tasks (comprising 31 fine-grained subtasks, 8 major categories, and over 10,000 manually verified instructions);
GeoDiffusion: A Training-Free Framework for Accurate 3D Geometric Conditioning in Image Generation
Phillip Mueller (University of Augsburg), Lars Mikelsons (University of Augsburg)
GenerationData SynthesisDiffusion modelImageBenchmark
🎯 What it does: Proposes an untrained GeoDiffusion framework that utilizes a single 3D reference model as a geometric prior to achieve precise image generation and editing under 3D geometric constraints.
GeoDistill: Geometry-Guided Self-Distillation for Weakly Supervised Cross-View Localization
Shaowen Tong (ShanghaiTech University), Yujiao Shi (ShanghaiTech University)
Pose EstimationAutonomous DrivingKnowledge DistillationImage
🎯 What it does: This paper proposes a self-distillation framework called GeoDistill based on FoV masking for weakly supervised cross-view localization.
GeoExplorer: Active Geo-localization with Curiosity-Driven Exploration
Li Mi (École Polytechnique Fédérale de Lausanne), Devis Tuia (École Polytechnique Fédérale de Lausanne)
Robotic IntelligenceTransformerReinforcement LearningVision-Language-Action ModelImageTextMultimodality
🎯 What it does: GeoExplorer is proposed, which combines goal-oriented and curiosity-driven rewards, utilizing causal Transformers for action-state dynamics modeling to achieve active geographic localization.
GeoFormer: Geometry Point Encoder for 3D Object Detection with Graph-based Transformer
Xin Jin (Chang'an University), Junchi Yan (Shanghai Jiao Tong University)
Object DetectionAutonomous DrivingGraph Neural NetworkTransformerPoint Cloud
🎯 What it does: This paper proposes GeoFormer, which constructs a geometric graph from point clouds within voxels and introduces graph weights in the Transformer to perform geometric point encoding for each voxel, achieving finer-grained voxel feature extraction.
GeoMan: Temporally Consistent Human Geometry Estimation using Image-to-Video Diffusion
Gwanghyun Kim (NVIDIA), Umar Iqbal (NVIDIA)
GenerationDepth EstimationDiffusion modelImageVideo
🎯 What it does: A geometric estimation method is proposed that can achieve temporal consistency while obtaining high-precision depth and normal maps using monocular human videos.
Geometric Alignment and Prior Modulation for View-Guided Point Cloud Completion on Unseen Categories
Jingqiao Xiu (National University of Singapore), Angela Yao (National University of Singapore)
Data SynthesisDepth EstimationTransformerPoint Cloud
🎯 What it does: A framework is proposed for view-guided point cloud completion that can operate under unseen categories, addressing the issue of traditional methods experiencing a sharp decline in performance across categories.
Geometry Distributions
Biao Zhang (King Abdullah University of Science and Technology), Peter Wonka (King Abdullah University of Science and Technology)
GenerationData SynthesisDiffusion modelPoint CloudMeshOrdinary Differential Equation
🎯 What it does: A geometric distribution representation method based on diffusion models is proposed, capable of mapping Gaussian noise to arbitrary topological, non-closed three-dimensional surface point sets, supporting infinite point sampling, texture encoding, and dynamic modeling.
GeometryCrafter: Consistent Geometry Estimation for Open-world Videos with Diffusion Priors
Tian-Xing Xu (Tsinghua University), Ying Shan (Tencent)
GenerationPose EstimationDepth EstimationDiffusion modelAuto EncoderVideoPoint Cloud
🎯 What it does: We propose GeometryCrafter, which can estimate temporally consistent, high-quality point maps from any open-world video for tasks such as 3D/4D reconstruction, camera pose estimation, depth editing, and video generation.
GEOPARD: Geometric Pretraining for Articulation Prediction in 3D Shapes
Pradyumn Goyal (University of Massachusetts Amherst), Evangelos Kalogerakis (University of Massachusetts Amherst)
SegmentationPose EstimationTransformerPoint Cloud
🎯 What it does: This work proposes a Transformer-based geometric pre-training framework called GEOPARD, which is used to predict the motion types, axes, and rotation centers of components from single-frame 3D shapes.
GeoProg3D: Compositional Visual Reasoning for City-Scale 3D Language Fields
Shunsuke Yasuki (Rikkyo University), Yutaka Matsuo (University of Tokyo)
Object DetectionSegmentationLarge Language ModelGaussian SplattingTextPoint CloudBenchmark
🎯 What it does: A GeoProg3D framework based on visual programming is proposed, enabling interpretable and accurate compositional geographic reasoning in high-fidelity 3D scenes at an urban scale through natural language.
GeoSplatting: Towards Geometry Guided Gaussian Splatting for Physically-based Inverse Rendering
Kai Ye (Peking University), Baoquan Chen (Peking University)
RestorationOptimizationGaussian SplattingPoint CloudMesh
🎯 What it does: This paper presents GeoSplatting, an inverse rendering method that combines 3D Gaussian projection (3DGS) with explicit mesh guidance, utilizing mesh surfaces to generate structured Gaussian points to improve normal estimation and light transport, thereby achieving efficient material-light decomposition.
GestureHYDRA: Semantic Co-speech Gesture Synthesis via Hybrid Modality Diffusion Transformer and Cascaded-Synchronized Retrieval-Augmented Generation
Quanwei Yang (University of Science and Technology of China), Hongtao Xie (University of Science and Technology of China)
GenerationData SynthesisTransformerDiffusion modelMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: A co-speech gesture synthesis system called GestureHYDRA based on a hybrid modal diffusion transformer has been developed, achieving explicit gesture generation driven by speech.
GestureLSM: Latent Shortcut based Co-Speech Gesture Generation with Spatial-Temporal Modeling
Pinxin Liu (University of Rochester), Chenliang Xu (University of Tokyo)
GenerationPose EstimationTransformerRectified FlowVideo
🎯 What it does: A full-body gesture generation model called GestureLSM is proposed, which can generate high-quality gestures in real-time based on flow matching and spatio-temporal attention.
GFPack++: Attention-Driven Gradient Fields for Optimizing 2D Irregular Packing
Tianyang Xue (Shandong University), Baoquan Chen (Peking University)
OptimizationGraph Neural NetworkTransformerScore-based ModelMeshStochastic Differential Equation
🎯 What it does: This paper proposes an attention-based extended version of GFPack++, which utilizes attention-encoded geometric and relational networks to learn gradient fields for efficient packing of irregular 2D shapes.
GGTalker: Talking Head Systhesis with Generalizable Gaussian Priors and Identity-Specific Adaptation
Wentao Hu (Beijing University of Posts and Telecommunications), Hui Tian (Beijing University of Posts and Telecommunications)
GenerationData SynthesisDiffusion modelGaussian SplattingVideoAudio
🎯 What it does: This paper presents GGTalker, a 3D talking head synthesis framework that utilizes a universal Gaussian prior and identity-specific adaptation, capable of achieving high-fidelity, fast audio-driven talking head animations from any viewpoint.
GigaTok: Scaling Visual Tokenizers to 3 Billion Parameters for Autoregressive Image Generation
Tianwei Xiong (University of Hong Kong), Xihui Liu (ByteDance)
GenerationTransformerContrastive LearningImage
🎯 What it does: This paper studies the contradiction between reconstruction and generation quality that arises when scaling the parameter size of visual tokenizers in visual autoregressive (AR) image generation. It proposes the GigaTok method, successfully scaling the visual tokenizer to 3 billion parameters while achieving the best performance in reconstruction, generation, and representation metrics on ImageNet.
GIViC: Generative Implicit Video Compression
Ge Gao (University of Bristol), David Bull (University of Bristol)
CompressionTransformerDiffusion modelVideo
🎯 What it does: This paper proposes GIViC, a full GOP-level video compression framework that utilizes implicit diffusion and hierarchical gated linear attention Transformer.
GlassWizard: Harvesting Diffusion Priors for Glass Surface Detection
Wenxue Li (Hong Kong University of Science and Technology), Lei Zhu (Hong Kong University of Science and Technology)
SegmentationDiffusion modelImageMultimodality
🎯 What it does: This paper addresses the task of glass surface detection (GSD) and proposes the GlassWizard framework, which utilizes a pre-trained diffusion model (Stable Diffusion) for glass segmentation.
GLEAM: Enhanced Transferable Adversarial Attacks for Vision-Language Pre-training Models via Global-Local Transformations
Yunqi Liu (Wuhan University), Xiaohui Cui (Wuhan University)
RetrievalAdversarial AttackTransformerVision Language ModelImageTextMultimodality
🎯 What it does: A unified framework GLEAM is designed and implemented to generate highly transferable adversarial samples for visual language pre-trained models in a black-box setting.
GLEAM: Learning Generalizable Exploration Policy for Active Mapping in Complex 3D Indoor Scene
Xiao Chen (Chinese University of Hong Kong), Tianfan Xue (Chinese University of Hong Kong)
Robotic IntelligenceTransformerReinforcement LearningPoint CloudBenchmark
🎯 What it does: A large-scale benchmark GLEAM-Bench and a unified RL exploration strategy GLEAM are proposed for active mapping in complex 3D indoor environments.
Global and Local Entailment Learning for Natural World Imagery
Srikumar Sastry (Washington University in St. Louis), Nathan Jacobs (Washington University in St. Louis)
ClassificationRetrievalTransformerContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a Radial Cross-Modal Embeddings (RCME) framework, which explicitly models hierarchical relationships through global entailment learning and hard negative sampling in visual-text models, achieving cross-modal alignment.
Global Motion Corresponder for 3D Point-Based Scene Interpolation under Large Motion
Junru Lin (University of Toronto), Ke Li (Simon Fraser University)
GenerationData SynthesisContrastive LearningPoint Cloud
🎯 What it does: This paper proposes the Global Motion Corresponder (GMC), which achieves 3D point cloud scene interpolation under large-scale motion between two frames by learning a unified potential field in SE(3) space.
Global Regulation and Excitation via Attention Tuning for Stereo Matching
Jiahao Li (City University of Hong Kong), Jianping Wang (Hon Hai Research Institute)
Depth EstimationAutonomous DrivingConvolutional Neural NetworkRecurrent Neural NetworkOptical FlowImage
🎯 What it does: A general framework called GREAT-Stereo is proposed, which integrates spatial, matching, and volumetric attention to improve iterative stereo matching.
Global-Aware Monocular Semantic Scene Completion with State Space Models
Shijie Li (A*STAR), Xulei Yang (A*STAR)
RestorationSegmentationTransformerImage
🎯 What it does: Proposes GA-MonoSSC, a hybrid architecture that combines Transformer and Mamba for monocular image completion and 3D semantic scene reconstruction.
GloPER: Unsupervised Animal Pattern Extraction from Local Reconstruction
Bowen Chen (University of Auckland), Gillian Dobbie (University of Auckland)
SegmentationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: An unsupervised animal texture extraction framework GloPER is proposed, which achieves fine-grained texture segmentation using local bi-color reconstruction;
GM-MoE: Low-Light Enhancement with Gated-Mechanism Mixture-of-Experts
Minwen Liao (Xinjiang University), Ziyang Yan (University of Trento)
RestorationConvolutional Neural NetworkMixture of ExpertsImage
🎯 What it does: The GM-MoE framework is proposed, utilizing a gated mixture of experts network to achieve comprehensive enhancement of low-light images.
GMMamba: Group Masking Mamba for Whole Slide Image Classification
Tingting Zheng (Harbin Institute of Technology), Sicheng Zhao (Tsinghua University)
ClassificationImage
🎯 What it does: In the whole slide image (WSI) classification task, the GMMamba framework is proposed, utilizing the group mask Mamba to perform local redundancy elimination on instances and enhancing global representation through cross-group super feature sampling.
Go to Zero: Towards Zero-shot Motion Generation with Million-scale Data
Ke Fan (Shanghai Jiao Tong University), Jingbo Wang (Shanghai AI Laboratory)
GenerationData SynthesisTransformerLarge Language ModelVideoTextBenchmark
🎯 What it does: A large-scale text-action pairing dataset called MotionMillion was constructed, and a scalable 7B parameter Transformer text-to-action generation model was trained based on this dataset. A zero-shot evaluation benchmark, MotionMillion-Eval, was introduced, demonstrating strong zero-shot generation capabilities.
Golden Noise for Diffusion Models: A Learning Framework
Zikai Zhou (Hong Kong University of Science and Technology Guangzhou), Zeke Xie (Hong Kong University of Science and Technology Guangzhou)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A noise prompting network NPNet is proposed and trained to convert random Gaussian noise into 'golden noise', thereby enhancing the image quality and semantic consistency of text-to-image diffusion models.
GRAB: A Challenging GRaph Analysis Benchmark for Large Multimodal Models
Jonathan Roberts (University of Cambridge), Samuel Albanie (University of Hong Kong)
Large Language ModelMultimodalityGraphBenchmark
🎯 What it does: This paper proposes a new large-scale multimodal model evaluation benchmark called GRAB, focusing on chart analysis tasks. It includes 3,284 high-quality questions synthesized using Matplotlib/Seaborn and provides 500 lightweight versions (GRAB-Lite) as well as a 'real' subset (hand-drawn, noisy, screenshots, etc.), aimed at challenging existing state-of-the-art models.
Gradient Decomposition and Alignment for Incremental Object Detection
Wenlong Luo (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
Object DetectionConvolutional Neural NetworkGaussian SplattingImage
🎯 What it does: A new incremental object detection framework GDA-IOD based on pseudo-labels and gradient alignment is proposed.
Gradient Extrapolation for Debiased Representation Learning
Ihab Asaad (Friedrich Schiller University Jena), Joachim Denzler (Friedrich Schiller University Jena)
Representation LearningImageText
🎯 What it does: This paper proposes the Gradient Extrapolation for Debiased Representation Learning (GERNE) method, which calculates gradients by sampling batches with varying degrees of bias and performs linear extrapolation to actively counteract spurious correlations during model training, thereby learning debiased representations.
Gradient Short-Circuit: Efficient Out-of-Distribution Detection via Feature Intervention
Jiawei Gu (Nanjing University of Science and Technology), Zechao Li (Nanjing University of Science and Technology)
Anomaly DetectionComputational EfficiencyImage
🎯 What it does: A method is proposed to identify and suppress feature dimensions that lead to excessively high confidence in OOD samples during the inference phase through Gradient Short-Circuit, thereby achieving efficient OOD detection.
Gradient-Reweighted Adversarial Camouflage for Physical Object Detection Evasion
Jiawei Liang (Shenzhen Campus of Sun Yat-sen University), Xiaochun Cao (Nanyang Technological University)
Object DetectionAutonomous DrivingAdversarial AttackImage
🎯 What it does: This paper proposes an adversarial camouflage method GRAC for physical object detection models, which can successfully induce misjudgments or missed detections by the detector under multiple perspectives and different lighting conditions.
Granular Concept Circuits: Toward a Fine-Grained Circuit Discovery for Concept Representations
Dahee Kwon (KAIST), Jaesik Choi (KAIST)
Explainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes the Granular Concept Circuit (GCC), which automatically identifies and constructs neuron circuits distributed across multiple layers to capture fine-grained concepts related to the query image in deep visual models.
Graph Domain Adaptation with Dual-branch Encoder and Two-level Alignment for Whole Slide Image-based Survival Prediction
Yuntao Shou (Xi'an Jiaotong University), Deyu Meng (Xi'an Jiaotong University)
Domain AdaptationGraph Neural NetworkImageBiomedical Data
🎯 What it does: Addressing the decline in survival prediction performance caused by domain transfer of WSI from different hospitals through a dual-branch graph encoder and a dual-layer alignment framework.
GraspCoT: Integrating Physical Property Reasoning for 6-DoF Grasping under Flexible Language Instructions
Xiaomeng Chu (University of Science and Technology of China), Yanyong Zhang (University of Science and Technology of China)
Robotic IntelligenceTransformerLarge Language ModelVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: The GraspCoT framework is proposed, achieving 6-DoF grasp detection based on physical properties through chain-of-thought (CoT) reasoning and multimodal LLM integration.
GReg: Geometry-Aware Region Refinement for Sign Language Video Generation
Tongkai Shi (Tianjin University), Wei Feng (Tianjin University)
GenerationData SynthesisGenerative Adversarial NetworkOptical FlowVideo
🎯 What it does: Designed and implemented the GReg framework, which utilizes the 3D geometric information of SMPL-X to refine regions in sign language videos, generating high-quality and temporally consistent sign language videos from source images.
Griffon v2: Advancing Multimodal Perception with High-Resolution Scaling and Visual-Language Co-Referring
Yufei Zhan (Chinese Academy of Sciences), Jinqiao Wang (Chinese Academy of Sciences)
RecognitionObject DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: A high-resolution multimodal model Griffon v2 is proposed, supporting visual inputs of up to 1K resolution and achieving visual and language co-reference capabilities.
GroundFlow: A Plug-in Module for Temporal Reasoning on 3D Point Cloud Sequential Grounding
Zijun Lin (Nanyang Technological University), Bihan Wen (Nanyang Technological University)
RecognitionObject DetectionComputational EfficiencyTransformerLarge Language ModelPoint CloudSequential
🎯 What it does: Designed and implemented the GroundFlow plugin module, injecting temporal reasoning capabilities into existing 3D visual grounding (3DVG) models, thereby efficiently completing the 3D point cloud sequence localization task (SG3D).
GroundingSuite: Measuring Complex Multi-Granular Pixel Grounding
Rui Hu (Huazhong University of Science and Technology), Xinggang Wang (Huazhong University of Science and Technology)
Object DetectionSegmentationTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper presents GroundingSuite, which includes the automatic annotation framework GSSculpt, a large-scale training set GSTrain-10M, and the evaluation benchmark GSEval, designed for pixel-level multi-granularity language-guided segmentation.
Group Inertial Poser: Multi-Person Pose and Global Translation from Sparse Inertial Sensors and Ultra-Wideband Ranging
Ying Xue (ETH Zurich), Christian Holz (ETH Zurich)
Pose EstimationOptimizationGraph Neural NetworkTime Series
🎯 What it does: Using sparse IMU and UWB distance data, a method for estimating multi-person body posture and global displacement (Group Inertial Poser) is proposed, along with a corresponding dataset.
Group-wise Scaling and Orthogonal Decomposition for Domain-Invariant Feature Extraction in Face Anti-Spoofing
Seungjin Jung (Chung Ang University), Jongwon Choi (Chung Ang University)
RecognitionDomain AdaptationContrastive LearningImage
🎯 What it does: The GD-FAS framework is proposed, which jointly aligns the weights and biases in facial anti-spoofing to enhance cross-domain generalization capabilities.
Grouped Speculative Decoding for Autoregressive Image Generation
Junhyuk So (POSTECH), Eunhyeok Park (POSTECH)
GenerationTransformerLarge Language ModelImage
🎯 What it does: Proposes the Grouped Speculative Decoding (GSD) method, which utilizes dynamically clustered visual token groups for untrained accelerated autoregressive image generation.
Growing a Twig to Accelerate Large Vision-Language Models
Zhenwei Shao (Zhejiang Key Laboratory of Space Information Sensing and Transmission), Jun Yu (Harbin Institute of Technology)
Computational EfficiencyTransformerVision Language ModelImageVideoTextMultimodality
🎯 What it does: Achieve inference acceleration on large visual language models by adding lightweight branch modules in the early layers.
GS-ID: Illumination Decomposition on Gaussian Splatting via Adaptive Light Aggregation and Diffusion-Guided Material Priors
Kang Du (Hong Kong University of Science and Technology), Zeyu Wang (Hong Kong University of Science and Technology)
RestorationGenerationOptimizationDiffusion modelGaussian SplattingImage
🎯 What it does: A lighting decomposition framework based on 3D Gaussian Splatting (GS-ID) is proposed, which can simultaneously separate geometry, material, and lighting from multi-view images under unknown lighting conditions, and supports editable local light sources and ambient light mixed rendering.
GS-LIVM: Real-Time Photo-Realistic LiDAR-Inertial-Visual Mapping with Gaussian Splatting
Yusen Xie (Hong Kong University of Science and Technology), Jun Ma (Hong Kong University of Science and Technology)
Autonomous DrivingOptimizationRobotic IntelligenceGaussian SplattingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper presents GS-LIVM, a real-time lighting realistic LiDAR-inertial-visual fusion SLAM framework that utilizes efficient 3D Gaussian splatting for instant rendering and map construction in large-scale outdoor scenes.
GS-Occ3D: Scaling Vision-only Occupancy Reconstruction with Gaussian Splatting
Baijun Ye (Tsinghua University), Hang Zhao (Tsinghua University)
SegmentationAutonomous DrivingGaussian SplattingPoint Cloud
🎯 What it does: A scalable visual-only occupancy reconstruction framework GS-Occ3D is proposed, utilizing 3D Gaussian splatting and Octree-based Gaussian surfels to achieve the separation and reconstruction of roads, backgrounds, and dynamic objects. Based on this, large-scale visual occupancy labels are automatically generated for training and enhancing 3D occupancy prediction models.
GSOT3D: Towards Generic 3D Single Object Tracking in the Wild
Yifan Jiao (Institute of Software Chinese Academy of Sciences), Libo Zhang (Institute of Software Chinese Academy of Sciences)
Object TrackingTransformerMultimodalityPoint CloudBenchmark
🎯 What it does: This paper presents the GSOT3D benchmark dataset and the PROT3D general 3D single-object tracking method, aiming to advance the research of 3D single-object tracking in outdoor environments.
GSRecon: Efficient Generalizable Gaussian Splatting for Surface Reconstruction from Sparse Views
Hang Yang (Nanjing University of Science and Technology), Jian Yang (Nanjing University)
Depth EstimationComputational EfficiencyGaussian SplattingPoint Cloud
🎯 What it does: A general sparse view surface reconstruction framework GSRecon based on efficient 3D Gaussian splatting is proposed.
GSV3D: Gaussian Splatting-based Geometric Distillation with Stable Video Diffusion for Single-Image 3D Object Generation
Ye Tao (Beihang University), Bin Zhou (Beihang University)
GenerationKnowledge DistillationTransformerDiffusion modelGaussian SplattingPoint CloudMesh
🎯 What it does: Using Stable Video Diffusion to generate multi-view latent variables, which are then transformed into renderable 3D representations through a Gaussian Splatting decoder, and incorporating geometric distillation (3D loss) during training to enhance multi-view consistency, achieving the generation of 3D objects from a single image.
GT-Loc: Unifying When and Where in Images Through a Joint Embedding Space
David G. Shatwell (University of Central Florida), Mubarak Shah (University of Central Florida)
RetrievalContrastive LearningImageMultimodality
🎯 What it does: This study proposes GT-Loc, a retrieval-based method that jointly predicts the time of image capture and geographic location.
GT-Mean Loss: A Simple Yet Effective Solution for Brightness Mismatch in Low-Light Image Enhancement
Jingxi Liao (Hefei University of Technology), Meng Wang (Hefei University of Technology)
RestorationTransformerImage
🎯 What it does: A GT-Mean loss function is proposed to address the issue of illumination mismatch in low-light image enhancement, improving the model's performance in brightness consistency and detail recovery.
GTR: Guided Thought Reinforcement Prevents Thought Collapse in RL-based VLM Agent Training
Tong Wei (Tsinghua University), Deheng Ye (Tencent)
Reinforcement LearningVision Language ModelImageChain-of-Thought
🎯 What it does: In the process of training VLM agents with RL, the GTR framework is introduced to automatically correct the thinking process, preventing thought collapse and improving the success rate of complex visual tasks.
GUAVA: Generalizable Upper Body 3D Gaussian Avatar
Dongbin Zhang (Tsinghua University), Haoqian Wang (Tsinghua University)
Image TranslationGenerationPose EstimationGaussian SplattingVideo
🎯 What it does: The GUAVA framework is proposed, which can quickly (≈0.1s) generate animatable upper-body 3D Gaussian avatars from a single image with pose information.
Guiding Diffusion Models with Adaptive Negative Sampling Without External Resources
Alakh Desai (University of California San Diego), Nuno Vasconcelos (University of California San Diego)
GenerationDiffusion modelImage
🎯 What it does: A training-free, resource-free adaptive negative sampling method named ANSWER is proposed, which improves the sampling process of diffusion models by dynamically estimating and applying negative guidance at each step, thereby enhancing the consistency between image quality and text prompts.
Guiding Diffusion-Based Articulated Object Generation by Partial Point Cloud Alignment and Physical Plausibility Constraints
Jens U. Kreber (University of Augsburg), Joerg Stueckler (University of Augsburg)
GenerationData SynthesisDiffusion modelPoint Cloud
🎯 What it does: This paper proposes PhysNAP, a framework for generating organic objects based on diffusion models, which can align using partial point clouds and enhance physical feasibility.
Guiding Noisy Label Conditional Diffusion Models with Score-based Discriminator Correction
Dat Nguyen Cong (FPT Software), Tung Hoang-Thanh (VNU University of Engineering and Technology)
GenerationData SynthesisDiffusion modelScore-based ModelImage
🎯 What it does: Without the need to retrain the generative model, the discriminator is used to correct the generation trajectory of the conditional diffusion model under noisy labels during the inference phase.
GUIOdyssey: A Comprehensive Dataset for Cross-App GUI Navigation on Mobile Devices
Quanfeng Lu (Shanghai AI Laboratory), Ping Luo (University of Hong Kong)
TransformerLarge Language ModelAgentic AIVision Language ModelMultimodality
🎯 What it does: This paper constructs a cross-application GUI navigation dataset called GUIOdyssey and designs OdysseyAgent—a multimodal navigation agent equipped with a historical resampling module based on this dataset.
GVDepth: Zero-Shot Monocular Depth Estimation for Ground Vehicles based on Probabilistic Cue Fusion
Karlo Koledić (University of Zagreb), Ivan Petrović (University of Zagreb)
Depth EstimationAutonomous DrivingConvolutional Neural NetworkImage
🎯 What it does: The GVDepth model is designed to achieve zero-shot monocular depth estimation by utilizing two geometric cues: vertical position and object size, through vertical normalization transformation and probabilistic weighted fusion.
GWM: Towards Scalable Gaussian World Models for Robotic Manipulation
Guanxing Lu (Tsinghua University), Siyuan Huang (State Key Laboratory of General Artificial Intelligence)
Robotic IntelligenceTransformerReinforcement LearningDiffusion modelAuto EncoderGaussian SplattingWorld ModelImageVideoPoint Cloud
🎯 What it does: Proposes the Gaussian World Model (GWM), which utilizes 3D Gaussian Splatting for explicit 3D modeling of scenes, encodes them into compact latent representations using a 3D Gaussian VAE, and employs a Diffusion Transformer (DiT) to predict future states under robotic action conditions for robot manipulation.
H3R: Hybrid Multi-view Correspondence for Generalizable 3D Reconstruction
Heng Jia (Singapore University of Technology and Design), Na Zhao (Zhejiang University)
GenerationDepth EstimationTransformerGaussian SplattingPoint Cloud
🎯 What it does: A hybrid multi-view correspondence framework named H3R is proposed, which utilizes voxel implicit fusion and camera-aware Transformer to directly generate high-quality 3D Gaussian representations in a single forward pass, supporting variable numbers of views and high-resolution inputs.
HADES: Human Avatar with Dynamic Explicit Hair Strands
Zhanfeng Liao (Tsinghua University), Yebin Liu (Tsinghua University)
GenerationData SynthesisGaussian SplattingVideo
🎯 What it does: Proposes the HADES framework, achieving seamless integration and animation of dynamic hair strands in full-body portraits.
HairCUP: Hair Compositional Universal Prior for 3D Gaussian Avatars
Byungjun Kim (Seoul National University), Junxuan Li (Codec Avatars Lab Meta)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A 3D Gaussian portrait universal prior model with separable face and hair is proposed, supporting seamless hairstyle changes and personalized avatar generation.
Hallucinatory Image Tokens: A Training-free EAZY Approach to Detecting and Mitigating Object Hallucinations in LVLMs
Liwei Che (Rutgers University), Vladimir Pavlovic (Rutgers University)
Object DetectionAnomaly DetectionTransformerVision Language ModelImageMultimodality
🎯 What it does: This paper analyzes the attention distribution of large visual-language models (LVLMs) and finds that a small number of high-attention image tokens (Hallucinatory Image Tokens, HITs) directly lead to the generation of hallucinated objects; it proposes a training-independent EAZY method that detects and eliminates hallucinations by zeroing out these HITs.
HAMoBE: Hierarchical and Adaptive Mixture of Biometric Experts for Video-based Person ReID
Yiyang Su (Michigan State University), Xiaoming Liu (Michigan State University)
RecognitionRetrievalTransformerMixture of ExpertsContrastive LearningVideo
🎯 What it does: This paper proposes a hierarchical and adaptive biometric expert mixture (HAMoBE) framework for video person re-identification, capable of extracting and fusing three types of key features: long-term, short-term, and temporal features from videos.
HAMSt3R: Human-Aware Multi-view Stereo 3D Reconstruction
Sara Rojas (King Abdullah University of Science and Technology), Grégory Rogez (NAVER LABS Europe)
SegmentationPose EstimationDepth EstimationKnowledge DistillationTransformerImagePoint Cloud
🎯 What it does: A full forward end-to-end multi-view sparse image 3D reconstruction model named HAMSt3R is proposed, capable of simultaneously reconstructing humans and scenes, outputting dense point clouds with human instance and DensePose information.
Harmonizing Visual Representations for Unified Multimodal Understanding and Generation
Size Wu (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
GenerationRepresentation LearningTransformerLarge Language ModelImageTextMultimodality
🎯 What it does: Proposes the Harmon framework, which uses a shared MAR encoder to simultaneously achieve visual understanding and image generation, forming a unified multimodal model.
HarmonySeg: Tubular Structure Segmentation with Deep-Shallow Feature Fusion and Growth-Suppression Balanced Loss
Yi Huang (Alibaba Group), Ke Yan (Alibaba Group)
SegmentationConvolutional Neural NetworkBiomedical DataMagnetic Resonance Imaging
🎯 What it does: The HarmonySeg framework is proposed, achieving precise segmentation of tubular structures such as blood vessels and airways in medical images through deep and shallow fusion, deep-to-shallow decoding, and growth suppression balanced loss.
Harnessing Input-Adaptive Inference for Efficient VLN
Dongwoo Kang (Oregon State University), Sanghyun Hong (Oregon State University)
Computational EfficiencyTransformerMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: A set of input-adaptive reasoning frameworks is proposed, which dynamically prunes unnecessary computations at runtime for the visual-language navigation (VLN) task, significantly reducing the computational power consumption of the visual encoder.
Harnessing Massive Satellite Imagery with Efficient Masked Image Modeling
Fengxiang Wang (National University of Defense Technology), Jing Zhang (Wuhan University)
ClassificationObject DetectionSegmentationTransformerImage
🎯 What it does: A 13M optical remote sensing image dataset, OpticalRS-13M, has been designed, and an efficient Masked Image Modeling pre-training method, SelectiveMAE, has been proposed for building remote sensing foundational models.
Harnessing Text-to-Image Diffusion Models for Point Cloud Self-Supervised Learning
Yiyang Chen (South China University of Technology), Dacheng Tao (Nanyang Technological University)
SegmentationGenerationRepresentation LearningTransformerDiffusion modelPoint Cloud
🎯 What it does: The PointSD framework is proposed, which replaces the Stable Diffusion text encoder with a 3D encoder. It first trains a diffusion model from point clouds to images, and then in the second stage aligns the intermediate features of SD with the features of the 3D backbone, achieving self-supervised pre-training of point clouds.
Harnessing Uncertainty-aware Bounding Boxes for Unsupervised 3D Object Detection
Ruiyang Zhang (University of Macau), Zhedong Zheng (University of Macau)
Object DetectionAutonomous DrivingPoint Cloud
🎯 What it does: This paper proposes a 3D object detection method called UA3D that does not require manual labels. By adding an auxiliary branch to the detection model, it estimates the uncertainty of each box coordinate by utilizing the differences in predicted box coordinates between the main and auxiliary branches. During training, the loss weights are adaptively adjusted based on this uncertainty to mitigate the impact of pseudo-label noise.
Harnessing Vision Foundation Models for High-Performance, Training-Free Open Vocabulary Segmentation
Yuheng Shi (University of Sydney), Chang Xu (University of Sydney)
SegmentationTransformerContrastive LearningImage
🎯 What it does: This paper proposes a training-free open-window word segmentation framework called Trident, which integrates CLIP, DINO, and SAM to achieve high-resolution semantic segmentation.
Hate in Plain Sight: On the Risks of Moderating AI-Generated Hateful Illusions
Yiting Qu (CISPA Helmholtz Center for Information Security), Yang Zhang (CISPA Helmholtz Center for Information Security)
GenerationData SynthesisAnomaly DetectionTransformerPrompt EngineeringVision Language ModelDiffusion modelImageText
🎯 What it does: Constructed and evaluated AI-generated 'hate illusions'—optical illusion images embedding hate speech or symbols into seemingly normal scenes, resulting in a dataset of 1,571 images;
HazeFlow: Revisit Haze Physical Model as ODE and Non-Homogeneous Haze Generation for Real-World Dehazing
Junseong Shin (Hanyang University), Tae Hyun Kim (Hanyang University)
RestorationData SynthesisRectified FlowImageOrdinary Differential Equation
🎯 What it does: A single image dehazing framework called HazeFlow based on ODE is proposed, which reinterprets the atmospheric scattering model as a learnable dynamic equation.
HccePose(BF): Predicting Front & Back Surfaces to Construct Ultra-Dense 2D-3D Correspondences for Pose Estimation
Yulin Wang (Southeast University), Chen Luo (Southeast University)
Pose EstimationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a method for simultaneously predicting the 3D coordinates of the front and back surfaces of a target and performing dense sampling between the two surfaces, constructing an ultra-dense 2D-3D correspondence to improve the accuracy of pose estimation based on PnP.
HDR Image Generation via Gain Map Decomposed Diffusion
Yuanshen Guan (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A framework for HDR image generation through diffusion models is proposed, which splits HDR images into SDR images and Gain Maps, and generates them jointly to obtain high dynamic range and wide color gamut images.
Head2Body: Body Pose Generation from Multi-sensory Head-mounted Inputs
Minh Tran (University of Southern California), Yelin Kim (Amazon)
GenerationPose EstimationTransformerVision Language ModelVideoMultimodality
🎯 What it does: The Head2Body framework is proposed, which infers full-body 3D posture using the IMU of a single head-mounted device and first-person visual input.