CVPR 2025 Papers — Page 11
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2871 papers
From Head to Tail: Efficient Black-box Model Inversion Attack via Long-tailed Learning
Ziang Li (Wuhan University), Chenjun Ma (Ant Group)
RecognitionAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a black-box model inversion attack based on long-tail enhancement (SMILE), achieving efficient reconstruction of private training data on high-resolution face recognition models.
From Head to Tail: Towards Balanced Representation in Large Vision-Language Models through Adaptive Data Calibration
Mingyang Song (Fudan University), Yu Cheng (Shanghai Artificial Intelligence Laboratory)
Large Language ModelVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: To address the long-tail distribution problem in instruction fine-tuning data for large-scale visual language models (LVLMs), this paper proposes an Adaptive Data Refinement (ADR) framework. It first filters and resamples the over-represented head concepts, and then generates visual and linguistic examples of tail concepts using text replacement and diffusion models, achieving data balance without increasing the data volume.
From Laboratory to Real World: A New Benchmark Towards Privacy-Preserved Visible-Infrared Person Re-Identification
Yan Jiang (Nanjing University of Information Science and Technology), Guoying Zhao (University of Oulu)
RecognitionFederated LearningSafty and PrivacyImageBenchmark
🎯 What it does: This paper proposes the L2RW benchmark, simulating privacy-preserving visible-infrared pedestrian re-identification scenarios in the real world.
From Multimodal LLMs to Generalist Embodied Agents: Methods and Lessons
Andrew Szot (Apple), Alexander Toshev (Apple)
Robotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision-Language-Action ModelMultimodality
🎯 What it does: This study explores how to transform multimodal large language models (MLLM) into general embodied agents (GEA) that can perform natural language instruction tasks across various domains such as control, gaming, UI, and planning through a unified action tokenizer.
From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization
Chao Yuan (Beihang University), Guanglin Niu (Beihang University)
RecognitionRetrievalDiffusion modelImage
🎯 What it does: A training-free portrait re-identification framework called Pose2ID is proposed, which suppresses noise and enhances identity representation by clustering features of the same identity.
From Prototypes to General Distributions: An Efficient Curriculum for Masked Image Modeling
Jinhong Lin (University of Wisconsin Madison), Pedro Morgado (University of Wisconsin Madison)
OptimizationRepresentation LearningTransformerAuto EncoderImage
🎯 What it does: This paper proposes a prototype-based curriculum learning framework to accelerate and enhance the pre-training effectiveness of Masked Image Modeling (MIM).
From Slow Bidirectional to Fast Autoregressive Video Diffusion Models
Tianwei Yin (Massachusetts Institute of Technology), Xun Huang (Adobe)
GenerationData SynthesisKnowledge DistillationTransformerDiffusion modelVideoTextOrdinary Differential Equation
🎯 What it does: Compress the traditional bidirectional diffusion model into a four-step causal generator to achieve low-latency real-time video generation.
From Sparse Signal to Smooth Motion: Real-Time Motion Generation with Rolling Prediction Models
German Barquero (Meta Reality Labs), Robin Kips (Meta Reality Labs)
GenerationPose EstimationRecurrent Neural NetworkVideoTime SeriesBenchmark
🎯 What it does: A Rolling Prediction Model (RPM) is proposed, which enables real-time generation of full-body movements, achieving smooth transitions even when hand tracking signals are missing or noisy, and provides the first paired VR hand/controller tracking with MoCap real data in the GOREP dataset.
From Sparse to Dense: Camera Relocalization with Scene-Specific Detector from Feature Gaussian Splatting
Zhiwei Huang (Zhejiang University), Guofeng Zhang (Zhejiang University)
Object DetectionPose EstimationGaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: This paper proposes an end-to-end sparse-to-dense camera relocalization pipeline called STDLoc, which utilizes Feature Gaussian as a scene representation to achieve accurate localization without prior pose information.
From Words to Structured Visuals: A Benchmark and Framework for Text-to-Diagram Generation and Editing
Jingxuan Wei (Shenyang Institute of Computing Technology Chinese Academy of Sciences), Ruifeng Guo (Shenyang Institute of Computing Technology Chinese Academy of Sciences)
GenerationTransformerLarge Language ModelImageTextBenchmark
🎯 What it does: This paper proposes the task of generating structured charts from text and provides corresponding benchmarks and an editing framework.
From Zero to Detail: Deconstructing Ultra-High-Definition Image Restoration from Progressive Spectral Perspective
Chen Zhao (Nanjing University), Ying Tai (Nanjing University)
RestorationSuper ResolutionTransformerImage
🎯 What it does: This paper proposes a step-by-step ultra-high-resolution image restoration framework ERR from a zero-to-detail spectral perspective, divided into three stages: zero-frequency enhancement, low-frequency recovery, and high-frequency refinement.
FrugalNeRF: Fast Convergence for Extreme Few-shot Novel View Synthesis without Learned Priors
Chin-Yang Lin (National Yang Ming Chiao Tung University), Yu-Lun Liu (NVIDIA Research)
GenerationData SynthesisComputational EfficiencyNeural Radiance FieldImage
🎯 What it does: Proposes FrugalNeRF, a neural radiance field framework that converges quickly with very few views;
FruitNinja: 3D Object Interior Texture Generation with Gaussian Splatting
Fangyu Wu (University of Waterloo), Yuhao Chen (University of Waterloo)
GenerationData SynthesisDiffusion modelGaussian SplattingImage
🎯 What it does: Generating internal textures for 3D objects and achieving real-time slicing rendering
FSBench: A Figure Skating Benchmark for Advancing Artistic Sports Understanding
Rong Gao (Lappeenranta-Lahti University of Technology), Heikki Kälviäinen (Lappeenranta-Lahti University of Technology)
ClassificationRecognitionSegmentationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVideoTextMultimodalityBenchmarkAudio
🎯 What it does: This paper presents the FSAnno dataset and the FSBench benchmark, constructing a multimodal, multi-level skating dataset that covers technical actions and artistic expressions, and designs a multi-task evaluation framework from prior knowledge to overall assessment.
FSboard: Over 3 Million Characters of ASL Fingerspelling Collected via Smartphones
Manfred Georg (Google), Thad Starner (Google)
RecognitionTransformerSupervised Fine-TuningVideoText
🎯 What it does: A large-scale American Sign Language finger-spelling dataset, FSboard, has been constructed, containing 266 hours and 3.2 million characters of video recorded by 147 deaf individuals using a Pixel 4A camera, along with a baseline model.
FSFM: A Generalizable Face Security Foundation Model via Self-Supervised Facial Representation Learning
Gaojian Wang (Zhejiang University), Kui Ren (Zhejiang University)
Representation LearningTransformerContrastive LearningImage
🎯 What it does: This paper proposes a self-supervised learning framework called FSFM, which learns a general representation of real human faces through two pre-training tasks: Masked Image Modeling and Instance Discrimination, and fine-tunes it for various face security tasks.
FSHNet: Fully Sparse Hybrid Network for 3D Object Detection
Shuai Liu (Sun Yat-sen University), Kai Huang (Sun Yat-sen University)
Object DetectionAutonomous DrivingTransformerPoint Cloud
🎯 What it does: This paper proposes a fully sparse 3D object detection network called FSHNet, which integrates sparse convolution with SlotFormer attention to achieve global interaction, and introduces dynamic sparse label assignment and sparse upsampling to enhance detection accuracy.
Full-DoF Egomotion Estimation for Event Cameras Using Geometric Solvers
Ji Zhao (Independent Researcher), Laurent Kneip (ShanghaiTech University)
Pose EstimationOptimizationSimultaneous Localization and MappingOptical FlowPoint Cloud
🎯 What it does: This paper proposes a sparse geometric solver that can fully recover the six degrees of freedom of camera motion (rotation and translation) using only the event stream generated by an event camera under the assumption of short time constant speed.
Functionality Understanding and Segmentation in 3D Scenes
Jaime Corsetti (Fondazione Bruno Kessler), Fabio Poiesi (Fondazione Bruno Kessler)
Object DetectionSegmentationTransformerLarge Language ModelVision Language ModelPoint CloudChain-of-Thought
🎯 What it does: Proposes Fun3DU, which utilizes pre-trained language models and vision-language models to achieve functional object segmentation in 3D scenes;
Fuzzy Multimodal Learning for Trusted Cross-modal Retrieval
Siyuan Duan (Sichuan University), Peng Hu (Sichuan University)
RetrievalContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a cross-modal retrieval method based on fuzzy set theory, called FUME, which can adaptively estimate and utilize uncertainty information to enhance retrieval credibility during the search process.
g3D-LF: Generalizable 3D-Language Feature Fields for Embodied Tasks
Zihan Wang (National University of Singapore), Gim Hee Lee (National University of Singapore)
Knowledge DistillationRobotic IntelligenceTransformerContrastive LearningTextPoint Cloud
🎯 What it does: This paper proposes g3D-LF, a 3D-language feature field model that can be constructed and updated in real-time in unseen environments.
G3Flow: Generative 3D Semantic Flow for Pose-aware and Generalizable Object Manipulation
Tianxing Chen (University of Hong Kong), Ping Luo (University of Hong Kong)
GenerationPose EstimationRobotic IntelligenceDiffusion modelPoint CloudBenchmark
🎯 What it does: Proposes the G3Flow framework, which constructs a 3D semantic flow in real-time for pose awareness and generalized object manipulation.
GA3CE: Unconstrained 3D Gaze Estimation with Gaze-Aware 3D Context Encoding
Yuki Kawana (Woven by Toyota), Norimasa Kobori (Woven by Toyota)
Pose EstimationTransformerImage
🎯 What it does: A 3D gaze estimation framework GA3CE based on 3D context encoding is proposed, which directly predicts the subject's 3D gaze direction from RGBD images and camera intrinsic parameters.
GAF: Gaussian Avatar Reconstruction from Monocular Videos via Multi-view Diffusion
Jiapeng Tang (Technical University of Munich), Matthias Nießner (Technical University of Munich)
GenerationData SynthesisDiffusion modelVideo
🎯 What it does: Using a multi-view diffusion model and 3D Gaussian light scattering technology, we reconstruct animatable photorealistic 3D Gaussian avatars of human faces from monocular videos.
Gain from Neighbors: Boosting Model Robustness in the Wild via Adversarial Perturbations Toward Neighboring Classes
Zhou Yang (Xidian University), Guangming Shi (Xidian University)
ClassificationObject DetectionTransformerImage
🎯 What it does: A robust training method GFN based on adjacent class gradient perturbation and inter-class distance weighted loss is proposed to enhance model performance under distribution shift.
Galaxy Walker: Geometry-aware VLMs For Galaxy-scale Understanding
Tianyu Chen (Beihang University), Jianxin Li (Beihang University)
ClassificationRecognitionGraph Neural NetworkTransformerMixture of ExpertsVision Language ModelImageMultimodality
🎯 What it does: We propose Galaxy-Walker, a visual-language model capable of multimodal learning of galaxy images and spectra in spherical, hyperbolic, and Euclidean geometric spaces;
GaPT-DAR: Category-level Garments Pose Tracking via Integrated 2D Deformation and 3D Reconstruction
Li Zhang (University of Science and Technology of China), Liu Liu (Hefei University of Technology)
Object TrackingPose EstimationDomain AdaptationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A category-level clothing pose tracking framework GaPT-DAR is proposed, based on 3D-2D projection, 2D deformation learning, and 3D depth reconstruction.
GarmentPile: Point-Level Visual Affordance Guided Retrieval and Adaptation for Cluttered Garments Manipulation
Ruihai Wu (Peking University), Hao Dong (Peking University)
RetrievalRobotic IntelligencePoint CloudBenchmark
🎯 What it does: This paper studies the task of clothing retrieval and reorganization in a chaotic pile of clothes, proposing a point cloud-based point-level visual empowerment model, and designing an adaptation module to reorganize the scene through pick-place actions when direct retrieval is not possible, facilitating subsequent retrieval.
GASP: Gaussian Avatars with Synthetic Priors
Jack Saunders (University of Bath), Benjamin E. Lundell (Microsoft)
GenerationData SynthesisAuto EncoderGaussian SplattingImageMesh
🎯 What it does: This paper proposes a Gaussian Avatar method based on synthetic priors (GASP), which can quickly construct animatable, panoramic-rendered realistic digital avatars using only a single camera or a single image.
GauCho: Gaussian Distributions with Cholesky Decomposition for Oriented Object Detection
José Henrique Lima Marques (Federal University of Rio Grande do Sul), Claudio R. Jung (Federal University of Rio Grande do Sul)
Object DetectionImage
🎯 What it does: A Gaussian distribution-based Cholesky decomposition regression head, GauCho, is proposed for directional object detection, directly regressing Gaussian parameters instead of traditional OBB parameters.
GaussHDR: High Dynamic Range Gaussian Splatting via Learning Unified 3D and 2D Local Tone Mapping
Jinfeng Liu (Hong Kong University of Science and Technology), Dan Xu (Hong Kong University of Science and Technology)
RestorationGenerationData SynthesisGaussian SplattingImage
🎯 What it does: We propose GaussHDR, a new HDR novel view synthesis method utilizing 3D Gaussian Splatting. By unifying the learning of 3D and 2D local tone mapping, it can reconstruct HDR scenes with high quality and generate LDR views that conform to exposure.
Gaussian Eigen Models for Human Heads
Wojciech Zielonka (Max Planck Institute for Intelligent Systems), Justus Thies (Max Planck Institute for Intelligent Systems)
GenerationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkGaussian SplattingImageVideo
🎯 What it does: This paper proposes a framework that distills high-quality 3D Gaussian avatars into a lightweight linear basis (GEM), capable of real-time expression prediction from a single RGB image and achieving cross-person expression transfer.
Gaussian Splashing: Unified Particles for Versatile Motion Synthesis and Rendering
Yutao Feng (Zhejiang University), Yin Yang (University of Utah)
GenerationData SynthesisComputational EfficiencyGaussian SplattingImage
🎯 What it does: A unified framework named Gaussian Splashing (GSP) is constructed, combining 3D Gaussian Splatting (3DGS) with Position-Based Dynamics (PBD) to achieve physical interactions and real-time rendering of solids, rigid bodies, and fluids in the same scene.
Gaussian Splatting Feature Fields for (Privacy-Preserving) Visual Localization
Maxime Pietrantoni (Czech Technical University in Prague), Torsten Sattler (Czech Technical University in Prague)
Pose EstimationSafty and PrivacyContrastive LearningGaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: This paper proposes Gaussian Splatting Feature Fields (GSFFs), a scene representation that combines 3D Gaussian Splatting with implicit feature fields for high-precision visual localization. It also designs an end-to-end self-supervised training framework that refines pose by aligning rendered 3D features with 2D image features, and utilizes clustering and segmentation to achieve a privacy-preserving localization pipeline.
Gaussian Splatting for Efficient Satellite Image Photogrammetry
Luca Savant Aira (Politecnico di Torino), Thibaud Ehret
GenerationComputational EfficiencyGaussian SplattingImage
🎯 What it does: A Earth Observation model based on 3D Gaussian splatting (EOGS) is proposed, specifically designed for digital surface modeling of multi-temporal satellite images.
GaussianFormer-2: Probabilistic Gaussian Superposition for Efficient 3D Occupancy Prediction
Yuanhui Huang (Tsinghua University), Jiwen Lu (Tsinghua University)
SegmentationAutonomous DrivingComputational EfficiencyTransformerGaussian SplattingPoint Cloud
🎯 What it does: This study proposes a 3D semantic occupancy prediction framework called GaussianFormer-2 based on probabilistic Gaussian superposition, improving the sparse Gaussian representation method of the original GaussianFormer.
GaussianIP: Identity-Preserving Realistic 3D Human Generation via Human-Centric Diffusion Prior
Zichen Tang (Beihang University), Hongyu Yang (Beihang University)
GenerationData SynthesisDiffusion modelGaussian SplattingImageText
🎯 What it does: A two-stage framework (GaussianIP) is proposed, capable of quickly generating identity-preserving, detail-rich facial features, and finely textured clothing for 3D human avatars based on text and portrait prompts.
GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting
Yangming Zhang (University of Texas at Arlington), Miao Yin (University of Texas at Arlington)
CompressionOptimizationGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes an optimization-based sparsification framework called GaussianSpa, which compresses 3D Gaussian Splatting (3DGS) models while maintaining high quality, significantly reducing the required number of Gaussian points.
GaussianUDF: Inferring Unsigned Distance Functions through 3D Gaussian Splatting
Shujuan Li (Tsinghua University), Zhizhong Han (Wayne State University)
RestorationGenerationOptimizationGaussian SplattingPoint CloudMesh
🎯 What it does: This paper proposes a UDF learning framework based on 3D Gaussian slicing plane projection to reconstruct high-quality surfaces with open boundaries from multi-view images.
GaussianWorld: Gaussian World Model for Streaming 3D Occupancy Prediction
Sicheng Zuo (Tsinghua University), Jiwen Lu (Tsinghua University)
SegmentationAutonomous DrivingConvolutional Neural NetworkWorld ModelPoint Cloud
🎯 What it does: This paper proposes the GaussianWorld framework based on a world model, transforming 3D occupancy prediction into 4D prediction, and utilizing historical 3D Gaussian representations for real-time occupancy inference in streaming scenes.
GaussTR: Foundation Model-Aligned Gaussian Transformer for Self-Supervised 3D Spatial Understanding
Haoyi Jiang (Huazhong University of Science and Technology), Xinggang Wang (Huazhong University of Science and Technology)
SegmentationAutonomous DrivingTransformerGaussian SplattingPoint Cloud
🎯 What it does: A Gaussian-based Transformer framework called GaussTR is proposed for self-supervised learning of 3D spatial perception, achieving zero-shot semantic occupancy prediction.
GauSTAR: Gaussian Surface Tracking and Reconstruction
Chengwei Zheng (ETH Zurich), Jie Song (Hong Kong University of Science and Technology)
RestorationObject TrackingGaussian SplattingOptical FlowVideo
🎯 What it does: This paper proposes the GauSTAR method, which utilizes 3D Gaussians bound to mesh surfaces to achieve realistic rendering, surface reconstruction, and 3D tracking of dynamic scenes, maintaining continuous tracking under topological changes (such as appearance/disappearance of surfaces, splitting/merging).
Gaze-LLE: Gaze Target Estimation via Large-Scale Learned Encoders
Fiona Ryan (Georgia Institute of Technology), James M. Rehg (University of Illinois Urbana-Champaign)
RecognitionObject DetectionTransformerImageVideo
🎯 What it does: Proposes Gaze-LLE, a single-stream gaze target prediction framework that uses only a frozen DINOv2 visual encoder along with a lightweight transformer decoder.
GazeGene: Large-scale Synthetic Gaze Dataset with 3D Eyeball Annotations
Yiwei Bao (Beihang University), Feng Lu (Beihang University)
GenerationData SynthesisPose EstimationImage
🎯 What it does: A large-scale synthetic gaze dataset called GazeGene is proposed, providing precise 3D eye structure annotations.
Gazing at Rewards: Eye Movements as a Lens into Human and AI Decision-Making in Hybrid Visual Foraging
Bo Wang (Nanyang Technological University), Mengmi Zhang (Nanyang Technological University)
TransformerReinforcement LearningImage
🎯 What it does: Developed and trained a Transformer-based Visual Forager (VF) for multi-object value mixed visual foraging tasks, and validated its performance through human eye-tracking experiments.
Gazing Into Missteps: Leveraging Eye-Gaze for Unsupervised Mistake Detection in Egocentric Videos of Skilled Human Activities
Michele Mazzamuto (University of Catania), Giovanni Maria Farinella (University of Catania)
Anomaly DetectionTransformerVideo
🎯 What it does: Unsupervised error detection using eye movement trajectories in first-person videos
GBC-Splat: Generalizable Gaussian-Based Clothed Human Digitalization under Sparse RGB Cameras
Hanzhang Tu (Tsinghua University), Yebin Liu (Tsinghua University)
SegmentationGenerationDepth EstimationGaussian SplattingImageMesh
🎯 What it does: Using sparse RGB camera input, we propose an end-to-end GBC-Splat pipeline that can quickly reconstruct complete human shape geometry and high-quality 2D Gaussian splatting representation from a few images, supporting real-time 6-DoF view rendering.
GBlobs: Explicit Local Structure via Gaussian Blobs for Improved Cross-Domain LiDAR-based 3D Object Detection
Dušan Malić (Graz University of Technology), Horst Possegger (Graz University of Technology)
Object DetectionDomain AdaptationAutonomous DrivingPoint Cloud
🎯 What it does: A local neighborhood encoding method based on Gaussian Blob is proposed for 3D object detection in LiDAR point clouds, significantly enhancing cross-domain generalization ability.
GCC: Generative Color Constancy via Diffusing a Color Checker
Chen-Wei Chang (National Yang Ming Chiao Tung University), Yu-Lun Liu (National Yang Ming Chiao Tung University)
RestorationGenerationDiffusion modelImage
🎯 What it does: By using diffusion models to generate and fill color checkers in images, a color-constant estimation of scene lighting is achieved.
GCE-Pose: Global Context Enhancement for Category-level Object Pose Estimation
Weihang Li (Technical University of Munich), Benjamin Busam (Technical University of Munich)
Pose EstimationImagePoint Cloud
🎯 What it does: For category-level object pose estimation, the GCE-Pose method is proposed, which first utilizes category priors to complete the semantic geometric reconstruction of partial RGB-D input (SSR module), and then fuses the globally reconstructed context with local features (GCE module), ultimately regressing the object's 6DoF pose and size.
GEAL: Generalizable 3D Affordance Learning with Cross-Modal Consistency
Dongyue Lu (National University of Singapore), Gim Hee Lee (National University of Singapore)
Robotic IntelligenceGaussian SplattingPoint Cloud
🎯 What it does: The GEAL framework is proposed, which maps point clouds to 2D using 3D Gaussian Splatting and leverages a pre-trained 2D model to enhance the generalization and robustness of 3D interaction space (affordance) learning.
GEM: A Generalizable Ego-Vision Multimodal World Model for Fine-Grained Ego-Motion, Object Dynamics, and Scene Composition Control
Mariam Hassan (École Polytechnique Fédérale de Lausanne), Alexandre Alahi (ETH Zurich)
GenerationPose EstimationAutonomous DrivingDiffusion modelWorld ModelVideoMultimodality
🎯 What it does: We propose GEM, a general multimodal self-perspective world model that enables controllable generation of videos from the perspectives of vehicles, drones, and humans, supporting target motion, perspective trajectories, and human pose control, while simultaneously outputting RGB and depth frames.
GEN3C: 3D-Informed World-Consistent Video Generation with Precise Camera Control
Xuanchi Ren, Jun Gao
GenerationData SynthesisAutonomous DrivingDiffusion modelVideoPoint Cloud
🎯 What it does: We propose GEN3C, a generative video model based on 3D caching that supports precise camera control and temporal consistency.
Gen3DEval: Using vLLMs for Automatic Evaluation of Generated 3D Objects
Shalini Maiti (Meta AI), Filippos Kokkinos (University College London)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelGaussian SplattingTextMeshBenchmark
🎯 What it does: An automatic evaluation framework Gen3DEval based on visual large language models (vLLM) is proposed and implemented to assess the appearance, surface quality, and text consistency of text-to-3D generation models.
GenAssets: Generating in-the-wild 3D Assets in Latent Space
Ze Yang (Waabi), Raquel Urtasun (Waabi)
GenerationData SynthesisAutonomous DrivingDiffusion modelPoint Cloud
🎯 What it does: On real-world sparse sensor data (cameras + LiDAR), we first learn a low-dimensional latent space of assets through occlusion-aware neural rendering, and then train a diffusion model on this latent space to generate complete and high-quality 3D traffic participant assets for multi-sensor simulation.
GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration
Sudarshan Rajagopalan (Johns Hopkins University), Vishal M. Patel (Johns Hopkins University)
RestorationData SynthesisTransformerDiffusion modelImage
🎯 What it does: Utilize diffusion models to generate various degraded images (fog, rain, snow, motion blur, low light, raindrops) to enhance image restoration training data across all scenarios.
Generalizable Object Keypoint Localization from Generative Priors
Dongkai Wang (Southwestern University of Finance and Economics), Shiliang Zhang (Peking University)
Object DetectionPose EstimationDiffusion modelImage
🎯 What it does: This paper proposes a general object keypoint localization method named GenLoc, which utilizes the prior knowledge from a pre-trained Latent Diffusion Model (LDM) generative model to achieve keypoint localization for multiple object categories without the need for large-scale labeled data.
Generalized Diffusion Detector: Mining Robust Features from Diffusion Models for Domain-Generalized Detection
Boyong He (Xiamen University), Liaoni Wu (Xiamen University)
Object DetectionDomain AdaptationKnowledge DistillationDiffusion modelImageBenchmark
🎯 What it does: This paper proposes a detection framework based on diffusion models, utilizing multi-step intermediate feature extraction during the diffusion process to achieve domain generalization detection.
Generalized Few-shot 3D Point Cloud Segmentation with Vision-Language Model
Zhaochong An (University of Copenhagen), Serge Belongie (University of Copenhagen)
SegmentationTransformerVision Language ModelPoint Cloud
🎯 What it does: A framework called GFS-VL is designed to achieve generalized few-shot 3D point cloud semantic segmentation by integrating a 3D visual language model with a small number of samples.
Generalized Gaussian Entropy Model for Point Cloud Attribute Compression with Dynamic Likelihood Intervals
Changhao Peng (Peking University)
CompressionAuto EncoderPoint Cloud
🎯 What it does: This paper proposes the use of a generalized Gaussian entropy model and dynamic likelihood intervals to improve probability estimation and arithmetic coding in point cloud attribute compression based on variational autoencoders.
Generalized Recorrupted-to-Recorrupted: Self-Supervised Learning Beyond Gaussian Noise
Brayan Monroy (Universidad Industrial de Santander), Julián Tachella (CNRS)
RestorationImage
🎯 What it does: This paper proposes a general self-supervised image restoration method GR2R, which can achieve unsupervised denoising under any natural exponential family noise.
Generalized Zero-Shot Classification via Semantics-Free Inter-Class Feature Generation
Libiao Chen (Beihang University), Zhenyu Tang (Beihang University)
ClassificationTransformerContrastive LearningBiomedical DataAlzheimer's Disease
🎯 What it does: A general zero-shot classification method that does not rely on semantic information is proposed, which sequentially models clinical risk levels by constructing a cross-class feature enhancement module to generate unseen class features.
Generalizing Deepfake Video Detection with Plug-and-Play: Video-Level Blending and Spatiotemporal Adapter Tuning
Zhiyuan Yan (Peking University), Li Yuan (Peking University)
ClassificationDomain AdaptationAnomaly DetectionConvolutional Neural NetworkGenerative Adversarial NetworkContrastive LearningVideo
🎯 What it does: This paper proposes video-level hybrid synthetic data (VB) to simulate facial feature drift (FFD) and introduces a lightweight spatiotemporal adapter (StA) to enhance the cross-dataset generalization performance of deepfake video detection models.
Generating 3D-Consistent Videos from Unposed Internet Photos
Gene Chou (Cornell University), Noah Snavely (Cornell University)
GenerationData SynthesisTransformerDiffusion modelGaussian SplattingImageVideo
🎯 What it does: This paper proposes a framework (KFC-W) that uses only 2-5 internet photos without pose as keyframes to generate coherent 3D perspective-consistent videos.
Generating 6DoF Object Manipulation Trajectories from Action Description in Egocentric Vision
Tomoya Yoshida (Kyoto University), Shinsuke Mori (Kyoto University)
GenerationRobotic IntelligenceTransformerLarge Language ModelVision Language ModelVideoPoint Cloud
🎯 What it does: The research automatically extracts and generates 6DoF object manipulation trajectories from first-person perspective videos, and builds a model to generate trajectories based on text descriptions.
Generating Multimodal Driving Scenes via Next-Scene Prediction
Yanhao Wu (Xining Jiao Tong University), Tong Zhang (École Polytechnique Fédérale de Lausanne)
GenerationData SynthesisAutonomous DrivingComputational EfficiencyTransformerGenerative Adversarial NetworkImageMultimodality
🎯 What it does: A unified multimodal driving scene generation framework (UMGen) is proposed, capable of simultaneously generating vehicle actions, road maps, traffic participants, and visual images in each frame, and supports users in customizing scenes based on the actions of the vehicle or other vehicles.
Generative Densification: Learning to Densify Gaussians for High-Fidelity Generalizable 3D Reconstruction
Seungtae Nam (Yonsei University), Eunbyung Park (Yonsei University)
RestorationGenerationData SynthesisTransformerGaussian SplattingPoint Cloud
🎯 What it does: A Generative Densification (GD) strategy is proposed, specifically designed for general feedforward Gaussian models (such as LaRa and MVSplat), which can selectively generate fine-grained Gaussian primitives for high-frequency detail areas in a single forward pass, thereby improving the quality of 3D reconstruction and viewpoint synthesis.
Generative Gaussian Splatting for Unbounded 3D City Generation
Haozhe Xie (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
GenerationData SynthesisComputational EfficiencyTransformerGenerative Adversarial NetworkGaussian SplattingPoint Cloud
🎯 What it does: A framework for unbounded city generation called GaussianCity is proposed, which utilizes a compact BEV-Point representation to achieve efficient unbounded city rendering.
Generative Hard Example Augmentation for Semantic Point Cloud Segmentation
Qi Zhang (Tianjin University), Di Lin (Tianjin University)
SegmentationGenerationAuto EncoderGenerative Adversarial NetworkPoint Cloud
🎯 What it does: This paper proposes the Generative Hard Example Augmentation (GHEA) method, which uses generative networks to generate variations in the latent space and reshape the source point cloud, producing diverse hard examples to enhance the training data for semantic point cloud segmentation.
Generative Image Layer Decomposition with Visual Effects
Jinrui Yang (University of California Santa Cruz), Yuyin Zhou (Adobe Research)
RestorationGenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: This paper proposes a layer decomposition framework called LAYERDECOMP based on diffusion Transformers, which can decompose input images into a clean background layer and a transparent foreground layer that retains visual effects such as shadows and reflections, supporting subsequent controllable layer editing.
Generative Inbetweening through Frame-wise Conditions-Driven Video Generation
Tianyi Zhu (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
GenerationData SynthesisPose EstimationDiffusion modelOptical FlowVideo
🎯 What it does: A frame-level conditional generative frame interpolation method called FCVG is proposed, which generates conditions for each frame by extracting matching lines and pose information, and injects them into the pre-trained Stable Video Diffusion model, significantly improving the temporal stability of video interpolation between key frames.
Generative Map Priors for Collaborative BEV Semantic Segmentation
Jiahui Fu (Beihang University), Si Liu (Beihang University)
SegmentationCompressionAutonomous DrivingDiffusion modelAuto EncoderImage
🎯 What it does: The CoGMP framework is proposed to achieve multi-vehicle collaborative BEV semantic segmentation, realizing efficient compression and robust fusion through generated map priors.
Generative Modeling of Class Probability for Multi-Modal Representation Learning
JungKyoo Shin (Chung Ang University), Eunwoo Kim (Chung Ang University)
RetrievalRepresentation LearningAuto EncoderContrastive LearningVideoTextMultimodality
🎯 What it does: A class anchor-based generative alignment method CALM is proposed for multimodal (video-text) representation learning.
Generative Multimodal Pretraining with Discrete Diffusion Timestep Tokens
Kaihang Pan (Zhejiang University), Hanwang Zhang (Nanyang Technological University)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: The paper proposes a visual language model based on Discrete Diffusion Time Steps (DDT), which encodes images as recursive discrete visual tokens and utilizes LLM to unify the next token prediction for multimodal understanding and generation tasks.
Generative Multiview Relighting for 3D Reconstruction under Extreme Illumination Variation
Hadi Alzayer (Google), Dor Verbin (Google)
RestorationGenerationDiffusion modelNeural Radiance FieldImagePoint Cloud
🎯 What it does: This paper proposes a method that first utilizes a multi-view diffusion model to unify the relighting of multiple images under different lighting conditions, and then employs NeRF-Casting with view-dependent shadow embedding for geometric and material reconstruction, achieving high-fidelity appearance reconstruction under extreme lighting variations.
Generative Omnimatte: Learning to Decompose Video into Layers
Yao-Chih Lee (Google DeepMind), Forrester Cole (Google DeepMind)
RestorationGenerationDiffusion modelVideo
🎯 What it does: Generate RGBA layer decomposition for videos (omnimatte) and achieve the removal and reconstruction of objects and their related effects such as shadows and reflections.
Generative Photography: Scene-Consistent Camera Control for Realistic Text-to-Image Synthesis
Yu Yuan (Purdue University), Stanley Chan (Purdue University)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: This paper proposes a generative photography framework that allows for precise control of camera intrinsic parameters (focal length, exposure, color temperature, bokeh, etc.) in text-to-image generation while maintaining scene consistency.
Generative Photomontage
Sean J. Liu (Carnegie Mellon University), Jun-Yan Zhu (Carnegie Mellon University)
SegmentationGenerationDiffusion modelImage
🎯 What it does: A generative collage framework is proposed, allowing users to select regions from multiple images generated by ControlNet using a brush, and then stitch these regions together to create the final image.
Generative Sparse-View Gaussian Splatting
Hanyang Kong (National University of Singapore), Xinchao Wang (National University of Singapore)
RestorationGenerationDepth EstimationDiffusion modelGaussian SplattingImageVideo
🎯 What it does: A Gaussian Splatting scheme (GS-GS) assisted by a generative diffusion model is proposed for high-quality 3D/4D scene reconstruction from sparse camera perspectives.
Generative Video Propagation
Shaoteng Liu (Chinese University of Hong Kong), Jiaya Jia (Hong Kong University of Science and Technology)
GenerationData SynthesisDiffusion modelVideo
🎯 What it does: A general generative video propagation framework called GenProp is designed, which can seamlessly propagate arbitrary edits from the first frame to the entire video, enabling various video editing tasks such as object removal, replacement, insertion, and tracking.
Generative Zero-Shot Composed Image Retrieval
Lan Wang (Michigan State University), Ser-Nam Lim (University of Central Florida)
GenerationRetrievalDiffusion modelImageText
🎯 What it does: This paper proposes a generative zero-shot composite image retrieval method called CIG, which first maps reference images to a text embedding space using a text inversion network, and then generates pseudo-target images using a latent diffusion model, thereby providing richer visual information for retrieval.
GenFusion: Closing the Loop between Reconstruction and Generation via Videos
Sibo Wu (Westlake University), Anpei Chen (University of Tübingen)
RestorationGenerationDiffusion modelVideo
🎯 What it does: This paper proposes a reconstruction-driven generation framework based on video diffusion models, utilizing video rendering for flawless 3D scene reconstruction and viewpoint synthesis.
GENIUS: A Generative Framework for Universal Multimodal Search
Sungyeon Kim (Amazon), Suha Kwak (POSTECH)
RetrievalTransformerVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes GENIUS, a universal generative retrieval framework that can automatically generate target IDs based on multimodal queries and instructions.
GENMANIP: LLM-driven Simulation for Generalizable Instruction-Following Manipulation
Ning Gao (Xi'an Jiaotong University), Jiangmiao Pang (Xi'an Jiaotong University)
Robotic IntelligenceTransformerLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: The GENMANIP platform and GENMANIP-BENCH benchmark are proposed, which use LLM to automatically generate task scenarios (Task-Oriented Scene Graphs, ToSG) and implement large-scale, diverse, and reproducible robotic manipulation experiments on IsaacSim, evaluating the performance of modular and end-to-end strategies in instruction following and scene generalization.
GenPC: Zero-shot Point Cloud Completion via 3D Generative Priors
An Li (Nanjing University of Aeronautics and Astronautics), Mingqiang Wei (Nanjing University of Aeronautics and Astronautics)
RestorationGenerationData SynthesisDepth EstimationAutonomous DrivingDiffusion modelPoint Cloud
🎯 What it does: A zero-shot point cloud completion framework called GenPC is proposed, which utilizes a pre-trained 3D generative model to complete missing point clouds from real scans.
GenVDM: Generating Vector Displacement Maps From a Single Image
Yuezhi Yang (University of Texas at Austin), Zhiqin Chen (Adobe Research)
GenerationData SynthesisDepth EstimationDiffusion modelImageMesh
🎯 What it does: A complete method for generating vector displacement maps (VDM) from a single image is proposed, which first utilizes a fine-tuned image diffusion model to generate multi-view normal maps, and then reconstructs the VDM through neural SDF and neural deformation fields to achieve seamless overlay of detail textures.
GeoAvatar: Geometrically-Consistent Multi-Person Avatar Reconstruction from Sparse Multi-View Videos
Soohyun Lee (Sogang University), Joo Ho Lee (Sogang University)
RestorationSegmentationPose EstimationGaussian SplattingVideo
🎯 What it does: For multi-person scenes in sparse multi-view videos, a multi-body head reconstruction method based on planar Gaussian splats is proposed, utilizing monocular priors and surface sorting techniques to address issues of penetration and occlusion in multi-body interactions.
GeoDepth: From Point-to-Depth to Plane-to-Depth Modeling for Self-Supervised Monocular Depth Estimation
Haifeng Wu (University of Electronic Science and Technology of China), Wen Li (University of Electronic Science and Technology of China)
Depth EstimationConvolutional Neural NetworkSupervised Fine-TuningImageVideo
🎯 What it does: Developed the GeoDepth framework, which utilizes a patch-to-depth modeling approach to directly predict plane normal vectors and offsets in self-supervised monocular depth estimation, rather than predicting depth pixel by pixel; achieved more continuous and accurate depth maps through a structured plane generation module and a depth discontinuity perception module.
Geometric Knowledge-Guided Localized Global Distribution Alignment for Federated Learning
Yanbiao Ma (Xidian University), Jiayi Chen (Xidian University)
Domain AdaptationFederated LearningImage
🎯 What it does: In federated learning, the GGEUR (Global Geometry-Guided Embedding Uncertainty Representation) method is proposed, which utilizes the geometric shape of the global embedding distribution to guide local data augmentation, thereby simulating the ideal global distribution on the client side.
Geometry Field Splatting with Gaussian Surfels
Kaiwen Jiang (University of California), Ravi Ramamoorthi (University of California)
RestorationGenerationGaussian SplattingImage
🎯 What it does: A geometric field splitting method based on Gaussian surfels is proposed to accurately reconstruct opaque surfaces from a set of calibrated RGB images, utilizing nearly exact differentiable rendering to enhance geometric quality.
Geometry in Style: 3D Stylization via Surface Normal Deformation
Nam Anh Dinh (University of Chicago), Rana Hanocka (University of Chicago)
GenerationOptimizationDiffusion modelMesh
🎯 What it does: Using text prompts to optimize target normal vectors of triangular meshes through a differentiable ARAP layer, achieving shape-preserving 3D stylization.
Geometry-guided Online 3D Video Synthesis with Multi-View Temporal Consistency
Hyunho Ha (Korea Advanced Institute of Science and Technology), Numair Khan (Meta)
GenerationData SynthesisConvolutional Neural NetworkGaussian SplattingVideo
🎯 What it does: This paper proposes an online multi-view video synthesis framework that utilizes global TSDF geometry guidance, 3D Gaussian forward rendering, and a CNN mixed network to achieve high-quality 3D video generation with consistent viewpoints and timing.
GeoMM: On Geodesic Perspective for Multi-modal Learning
Shibin Mei (Huawei), Bingbing Ni (Shanghai Jiao Tong University)
RetrievalRepresentation LearningTransformerContrastive LearningImageTextMultimodality
🎯 What it does: This paper introduces geodesic distance as a new distance metric within the contrastive learning framework of multimodal learning, and proposes an efficient computation and incremental update algorithm based on hierarchical graphs, enabling quick and accurate retrieval of geodesic distances in large-scale feature queues. Subsequently, this distance is integrated into the contrastive loss of existing models such as CLIP and ALBEF, followed by pre-training and evaluation across multiple tasks.
Ges3ViG : Incorporating Pointing Gestures into Language-Based 3D Visual Grounding for Embodied Reference Understanding
Atharv Mahesh Mane (Stony Brook University), Archan Misra (Massachusetts Institute of Technology)
Object DetectionSegmentationData SynthesisPose EstimationTransformerVision Language ModelTextPoint Cloud
🎯 What it does: This paper proposes an enhanced framework called Imputer that automates the integration of human pointing gestures with language instructions into 3D scenes, and utilizes it to generate a new 3D-ERU dataset named ImputeRefer, while designing a unified model called Ges3ViG to achieve human localization, gesture understanding, and language fusion.
GET: Unlocking the Multi-modal Potential of CLIP for Generalized Category Discovery
Enguang Wang (Nankai University), Ming-Ming Cheng (Nankai University)
ClassificationKnowledge DistillationRepresentation LearningContrastive LearningImageMultimodality
🎯 What it does: Utilize CLIP multimodal features to achieve joint discovery of known and unknown categories in unlabeled data.
GFlowVLM: Enhancing Multi-step Reasoning in Vision-Language Models with Generative Flow Networks
Haoqiang Kang (Honda Research Institute USA), Kwonjoon Lee (Honda Research Institute USA)
Robotic IntelligenceTransformerReinforcement LearningVision Language ModelFlow-based ModelMultimodalityChain-of-Thought
🎯 What it does: Fine-tuning the visual language model (VLM) for multi-step reasoning tasks using Generative Flow Networks (GFlowNet) to enhance its diversity and generalization ability in long sequence decision-making.
GG-SSMs: Graph-Generating State Space Models
Nikola Zubic (University of Zurich), Davide Scaramuzza (University of Zurich)
Autonomous DrivingOptimizationGraph Neural NetworkOptical FlowImageGraphTime Series
🎯 What it does: A graph generation state space model (GG-SSM) is proposed, which captures long-range dependencies in high-dimensional data by dynamically constructing a minimum spanning tree graph.
GIF: Generative Inspiration for Face Recognition at Scale
Saeed Ebrahimi (West Virginia University), Nasser M. Nasrabadi (West Virginia University)
RecognitionComputational EfficiencyConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Proposes the GIF framework, which uses structured identity codes to replace traditional scalar labels for training facial recognition models, thereby reducing the Softmax computational cost from O(m) to O(log m).
GIFStream: 4D Gaussian-based Immersive Video with Feature Stream
Hao Li (Zhejiang University), Yiyi Liao (Zhejiang University)
GenerationCompressionGaussian SplattingVideo
🎯 What it does: This paper proposes GIFStream, a 4D Gaussian representation that incorporates time-related feature streams based on deformable 3D Gaussians, achieving high-quality rendering and low storage for immersive videos.
GigaHands: A Massive Annotated Dataset of Bimanual Hand Activities
Rao Fu (Brown University), Srinath Sridhar (Brown University)
Object DetectionObject TrackingSegmentationData SynthesisPose EstimationLarge Language ModelImageVideoText
🎯 What it does: The GigaHands dataset is proposed and released, recording 34 hours of dual hand activities, covering 417 types of objects and 56 subjects, including 183M frames of RGB, 3.7M 3D gestures, and 84k text descriptions.