ICCV 2025 Papers — Page 10
IEEE/CVF International Conference on Computer Vision · 2701 papers
Forgetting Through Transforming: Enabling Federated Unlearning via Class-Aware Representation Transformation
Qi Guo (Xi'an Jiaotong University), Bingyi Liu (Wuhan University of Technology)
ClassificationFederated LearningImage
🎯 What it does: A federated unlearning framework FUCRT based on category-aware representation transformation is proposed, achieving efficient forgetting of specified category data in a federated learning environment.
FoundIR: Unleashing Million-scale Training Data to Advance Foundation Models for Image Restoration
Hao Li (Nanjing University of Science and Technology), Jinshan Pan (Nanjing University of Science and Technology)
RestorationDiffusion modelImage
🎯 What it does: A dataset of over one million high-quality real aligned image pairs was constructed, and a general image restoration model called FoundIR was proposed, which combines a diffusion general model with a degradation-aware expert model, using an incremental learning training strategy.
FPEM: Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching
Hui Li (Alibaba Group), Xu Liu (Shanghai Jiao Tong University)
ClassificationRecognitionTransformerContrastive LearningImageVideoMultimodality
🎯 What it does: A real-time attractiveness prediction framework for facial aesthetics in live video is proposed, and a large-scale dataset LiveBeauty consisting of 10,000 live facial images and 200,000 ratings is constructed.
FrameFusion: Combining Similarity and Importance for Video Token Reduction on Large Vision Language Models
Tianyu Fu (Tsinghua University), Yu Wang (Infinigence AI)
CompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideo
🎯 What it does: A video visual token compression method named FrameFusion is proposed, which significantly reduces the number of visual tokens by combining similarity merging and importance pruning.
FramePainter: Endowing Interactive Image Editing with Video Diffusion Priors
Yabo Zhang (Harbin Institute of Technology), Wangmeng Zuo (Huawei Noah's Ark Lab)
GenerationData SynthesisDiffusion modelOptical FlowImageVideo
🎯 What it does: This paper proposes FramePainter, a framework that reconstructs interactive image editing tasks as image-to-video generation tasks, utilizing video diffusion models for fine and controllable editing.
Free-Form Motion Control: Controlling the 6D Poses of Camera and Objects in Video Generation
Xincheng Shuai (Fudan University), Dacheng Tao (Nanyang Technological University)
GenerationData SynthesisPose EstimationDiffusion modelVideo
🎯 What it does: This paper presents the SynFMC dataset and the FMC method, achieving free control of the 6D poses of the camera and objects in video generation.
FREE-Merging: Fourier Transform for Efficient Model Merging
Shenghe Zheng (Harbin Institute of Technology), Hongzhi Wang (Harbin Institute of Technology)
OptimizationComputational EfficiencyTransformerMixture of ExpertsTextMultimodality
🎯 What it does: This paper proposes a model merging method based on frequency domain high-pass filtering, called FR-Merging, and a lightweight expert module, named FREE-Merging, which can merge multi-task models without additional training, balancing storage and inference efficiency.
Free-MoRef: Instantly Multiplexing Context Perception Capabilities of Video-MLLMs within Single Inference
Kuo Wang (Sun Yat-sen University), Guanbin Li (Sun Yat-sen University)
RecognitionGenerationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsVision Language ModelVideo
🎯 What it does: Proposes the Free-MoRef method, which achieves long context awareness in video LLMs through multi-reference splitting and MoRef attention in a single inference.
Free-running vs Synchronous: Single-Photon Lidar for High-flux 3D Imaging
Ruangrawee Kitichotkul (Boston University), Vivek K Goyal (Mitsubishi Electric Research Laboratories)
Depth EstimationOptimizationDiffusion modelScore-based ModelPoint CloudMesh
🎯 What it does: This study compares the performance of single-photon radar in free-running and synchronized modes, proposing a joint maximum likelihood estimator and a depth regularization algorithm based on point cloud scoring.
Free2Guide: Training-Free Text-to-Video Alignment using Image LVLM
Jaemin Kim (KAIST), Jong Chul Ye (KAIST)
GenerationData SynthesisVision Language ModelDiffusion modelVideoText
🎯 What it does: This paper proposes Free Guide 2, a training-independent and gradient-independent text-to-video alignment framework that utilizes a large visual language model (LVLM) at the image level as a reward function, and guides the diffusion sampling process through path integral control.
Free4D: Tuning-free 4D Scene Generation with Spatial-Temporal Consistency
Tianqi Liu (Huazhong University of Science and Technology), Ziwei Liu (Nanyang Technological University)
GenerationData SynthesisDiffusion modelGaussian SplattingImageVideoPoint CloudBenchmark
🎯 What it does: This paper presents Free4D, a framework that generates complete 4D scenes (space + time) from a single image without the need for parameter tuning, capable of rendering high-quality, coherent dynamic videos from any viewpoint.
FreeCus: Free Lunch Subject-driven Customization in Diffusion Transformers
Yanbing Zhang (East China University of Science and Technology), Mengping Yang (Shanghai Academy of Artificial Intelligence)
GenerationTransformerLarge Language ModelDiffusion modelTextMultimodality
🎯 What it does: Developed the FreeCus framework, achieving zero-shot, no-training topic-customized text generation that supports diverse contexts and styles while maintaining subject consistency;
FreeDance: Towards Harmonic Free-Number Group Dance Generation via a Unified Framework
Yiwen Zhao (Carnegie Mellon University), Xingqun Qi (Peking University)
GenerationTransformerAuto EncoderVideoMultimodality
🎯 What it does: A unified framework called FreeDance is proposed to generate group dances for any number of participants under musical conditions.
FreeDNA: Endowing Domain Adaptation of Diffusion-Based Dense Prediction with Training-Free Domain Noise Alignment
Hang Xu (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
SegmentationDepth EstimationSuper ResolutionDomain AdaptationDiffusion modelOptical FlowImage
🎯 What it does: A domain noise alignment (DNA) method is proposed that does not require training, utilizing the alignment of noise statistics from diffusion models to achieve domain adaptation for dense prediction tasks, accommodating both scenarios where the source domain is available and unavailable.
FreeFlux: Understanding and Exploiting Layer-Specific Roles in RoPE-Based MMDiT for Versatile Image Editing
Tianyi Wei (Nanyang Technological University), Xingang Pan (Nanyang Technological University)
GenerationTransformerDiffusion modelImageTextMultimodality
🎯 What it does: A mechanism analysis of the RoPE-encoded multimodal diffusion transformer (MMDiT) model (FLUX) is conducted, and a training-free multi-task image editing framework is proposed based on the hierarchical dependency between position encoding and content similarity. This framework adaptively injects key/value vectors according to the characteristics of the editing tasks (position-related, content-similar, region-preserving) and introduces an inference-first generation and value replacement strategy.
FreeMorph: Tuning-Free Generalized Image Morphing with Diffusion Model
Yukang Cao (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: We propose FreeMorph, a completely non-tunable image morphing method that can generate a sequence of intermediate frames smoothly transitioning from one image to another within 30 seconds.
FreeScale: Unleashing the Resolution of Diffusion Models via Tuning-Free Scale Fusion
Haonan Qiu (Nanyang Technological University), Ziwei Liu (Fudan University)
GenerationData SynthesisDiffusion modelImageVideo
🎯 What it does: This paper presents FreeScale, a parameter-free inference paradigm that achieves high-resolution (up to 8K) image and video generation from pre-trained diffusion models through multi-scale fusion and frequency domain extraction.
FreeSplatter: Pose-free Gaussian Splatting for Sparse-view 3D Reconstruction
Jiale Xu (Tencent), Ying Shan (Tencent)
Pose EstimationTransformerGaussian SplattingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes a posture-free sparse view 3D reconstruction framework called FreeSplatter, which can generate high-quality 3D Gaussian point clouds directly through a unified transformer without calibrated camera parameters and estimate camera poses within seconds.
FreqPDE: Rethinking Positional Depth Embedding for Multi-View 3D Object Detection Transformers
Haisheng Su (Shanghai Jiao Tong University), Nanning Zheng (Xi'an Jiaotong University)
Object DetectionDepth EstimationAutonomous DrivingConvolutional Neural NetworkTransformerImagePoint Cloud
🎯 What it does: A multi-view 3D detection method based on frequency domain information, FreqPDE, is proposed, which enhances the depth representation and localization accuracy of 3D detection using a frequency-aware spatial pyramid encoder, cross-view scale-invariant depth predictor, and position depth encoder.
Frequency Domain-Based Diffusion Model for Unpaired Image Dehazing
Chengxu Liu (Xi'an Jiaotong University), Ming-Hsuan Yang (University of California)
Image TranslationRestorationDiffusion modelContrastive LearningImage
🎯 What it does: A frequency domain-based diffusion model (FrDiff) is proposed for unpaired image dehazing, achieving frequency domain reconstruction and correction of haze loss through an Amplitude Residual Encoder (ARE) and a Phase Correction Module (PCM).
Frequency-Aligned Knowledge Distillation for Lightweight Spatiotemporal Forecasting
Yuqi Li (Institute of Computing Technology Chinese Academy of Sciences), Hao Wu (Tsinghua University)
Knowledge DistillationConvolutional Neural NetworkTransformerTime SeriesSequential
🎯 What it does: A frequency-aligned knowledge distillation framework SDKD is proposed for lightweight spatiotemporal prediction, which transfers multi-scale spatiotemporal features to the student network through high-frequency convolution and low-frequency Transformer decomposition of the teacher model.
Frequency-Aware Autoregressive Modeling for Efficient High-Resolution Image Synthesis
Zhuokun Chen (South China University of Technology), Mingkui Tan (South China University of Technology)
GenerationData SynthesisComputational EfficiencyTransformerImage
🎯 What it does: Proposes the SparseVAR framework, dynamically removing low-frequency tokens during the high-resolution phase to accelerate visual autoregressive models.
Frequency-Dynamic Attention Modulation For Dense Prediction
Linwei Chen (Beijing Institute of Technology), Ying Fu (University of Tokyo)
Object DetectionSegmentationTransformerImage
🎯 What it does: Proposes Frequency-Dynamic Attention Modulation (FDAM), which enhances the spectral representation of ViT through Attention Inversion (AttInv) and Frequency Dynamic Scaling (FreqScale), addressing the frequency disappearance issue caused by low-pass filtering.
Frequency-Guided Diffusion for Training-Free Text-Driven Image Translation
Zheng Gao (Queen Mary University of London), Ioannis Patras (Queen Mary University of London)
Image TranslationDiffusion modelImage
🎯 What it does: A frequency-guided diffusion model (FGD) is proposed, which utilizes low-frequency/high-frequency information from the source image to guide text-driven image translation, achieving preservation and alteration of style or structure.
Frequency-Guided Posterior Sampling for Diffusion-Based Image Restoration
Darshan Thaker (University of Pennsylvania), Rene Vidal
RestorationDiffusion modelImage
🎯 What it does: A frequency-guided posterior sampling (FGPS) method is proposed, which gradually incorporates frequency information into the reverse diffusion process using a time-varying low-pass filter to address the approximation errors of traditional diffusion recovery methods.
Frequency-Semantic Enhanced Variational Autoencoder for Zero-Shot Skeleton-based Action Recognition
Wenhan Wu (University of North Carolina at Charlotte), Aidong Lu (University of North Carolina at Charlotte)
RecognitionPose EstimationGraph Neural NetworkLarge Language ModelAuto EncoderVideoText
🎯 What it does: The FS-VAE framework is proposed, which implements zero-shot skeleton action recognition through three main modules: frequency enhancement, semantic action description, and calibration alignment.
FRET: Feature Redundancy Elimination for Test Time Adaptation
Linjing You (Institute of Automation, Chinese Academy of Sciences), Xiangli Nie (Institute of Automation, Chinese Academy of Sciences)
Domain AdaptationGraph Neural NetworkContrastive LearningImageMultimodality
🎯 What it does: This paper proposes a method to enhance model generalization in Test-Time Adaptation (TTA) by eliminating feature redundancy, and implements two approaches—S-FRET (direct minimization of redundancy score) and G-FRET (a graph-based method combining GCN and contrastive learning).
From Abyssal Darkness to Blinding Glare: A Benchmark on Extreme Exposure Correction in Real World
Bo Wang (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)
RestorationDiffusion modelContrastive LearningImageBenchmark
🎯 What it does: The REED extreme exposure dataset is proposed, and the Context-Guided Luminance-Normalized Iterative Exposure Refinement Network (CLIER) is designed to restore extreme exposure images.
From Easy to Hard: Progressive Active Learning Framework for Infrared Small Target Detection with Single Point Supervision
Chuang Yu (Chinese Academy of Sciences), Xiangyu Yue (Chinese University of Hong Kong)
Object DetectionSupervised Fine-TuningImage
🎯 What it does: A progressive active learning framework (PAL) is proposed for single-point supervised infrared small target detection, ranging from easy to difficult;
From Easy to Hard: The MIR Benchmark for Progressive Interleaved Multi-Image Reasoning
Hang Du (Beijing University of Posts and Telecommunications), Sicong Leng
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningImageMultimodalityBenchmark
🎯 What it does: A multi-image interleaved reasoning benchmark (MIR) is proposed, along with a five-step reasoning process for each sample; a difficulty-based staged curriculum learning method is also introduced to gradually enhance the reasoning ability of multimodal large language models (MLLMs).
From Enhancement to Understanding: Build a Generalized Bridge for Low-light Vision via Semantically Consistent Unsupervised Fine-tuning
Sen Wang (East China Normal University), Lizhuang Ma (East China Normal University)
RestorationObject DetectionSegmentationDomain AdaptationDiffusion modelImage
🎯 What it does: By constructing an unsupervised refinement framework based on diffusion models, a method called 'Generalized Low-light Enhancement and Understanding (GEFU)' is proposed, which can enhance low-light images directly to a quality suitable for various visual tasks without the need for annotations.
From Gallery to Wrist: Realistic 3D Bracelet Insertion in Videos
Chenjian Gao (Chinese University of Hong Kong), Tianfan Xue (Chinese University of Hong Kong)
GenerationData SynthesisDiffusion modelGaussian SplattingImageVideo
🎯 What it does: A hybrid video object insertion pipeline is proposed that combines 3D Gaussian splatting rendering with a 2D diffusion model, achieving high temporal consistency and realistic lighting for dynamically worn wristbands.
From Gaze to Movement: Predicting Visual Attention for Autonomous Driving Human-Machine Interaction based on Programmatic Imitation Learning
Yexin Huang (Tongji University), Jie Wang (Tongji University)
Autonomous DrivingOptimizationExplainability and InterpretabilityVideo
🎯 What it does: Developed a PILOT model based on programmatic imitation learning to predict driver eye movement trajectories using rule expressions, and proposed the DATAD dataset for automated driving takeover scenarios.
From Holistic to Localized: Local Enhanced Adapters for Efficient Visual Instruction Fine-Tuning
Pengkun Jiao (Fudan University), Yu-Gang Jiang (Fudan University)
RecognitionOptimizationComputational EfficiencyTransformerSupervised Fine-TuningMixture of ExpertsImageMultimodality
🎯 What it does: Proposes two modules, Dual-LoRA and VCE, for efficient visual instruction fine-tuning, addressing data information conflicts in multi-task scenarios.
From Image to Video: An Empirical Study of Diffusion Representations
Pedro Vélez, Mehdi S. M. Sajjadi
ClassificationRecognitionObject TrackingGenerationPose EstimationDepth EstimationTransformerDiffusion modelImageVideo
🎯 What it does: Compare the representation capabilities of the same architecture (WALT) under pre-training for image generation and video generation, and evaluate its performance in various visual downstream tasks (image classification, action recognition, depth estimation, camera pose estimation, point/frame tracking).
From Imitation to Innovation: The Emergence of AI's Unique Artistic Styles and the Challenge of Copyright Protection
Zexi Jia (Tencent Inc), Jie Zhou (Tencent Inc)
ClassificationGenerationExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelContrastive LearningImageMultimodalityBenchmark
🎯 What it does: An interpretable AI art copyright determination framework called ArtBulb is proposed and implemented, along with the construction of the first AI art copyright benchmark dataset AICD.
From Linearity to Non-Linearity: How Masked Autoencoders Capture Spatial Correlations
Anthony Bisulco (University of Pennsylvania), Pratik Chaudhari (University of Pennsylvania)
Representation LearningHyperparameter SearchTransformerAuto EncoderImage
🎯 What it does: The theoretical and experimental analysis of the mechanism of Masked Autoencoders (MAE) in learning spatial correlations, along with practical suggestions for hyperparameter selection.
From Objects to Events: Unlocking Complex Visual Understanding in Object Detectors via LLM-guided Symbolic Reasoning
Yuhui Zeng (Xiamen University), Rongrong Ji (Tencent)
Object DetectionAnomaly DetectionExplainability and InterpretabilityLarge Language ModelPrompt EngineeringImageVideo
🎯 What it does: Proposes the SymbolicDet framework, which utilizes LLM-guided symbolic reasoning to transform the low-level object recognition capabilities of standard object detectors into complex event recognition without the need for training.
From One to More: Contextual Part Latents for 3D Generation
Shaocong Dong (Hong Kong University of Science and Technology), Dan Xu (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderPoint CloudMesh
🎯 What it does: A 3D generation framework called CoPart is proposed, which is based on contextual component latent variables and can decompose objects into multiple geometric + image latent variables and generate them synchronously.
From Panels to Prose: Generating Literary Narratives from Comics
Ragav Sachdeva (University of Oxford), Andrew Zisserman (University of Oxford)
Object DetectionGenerationTransformerLarge Language ModelPrompt EngineeringTextMultimodality
🎯 What it does: This paper proposes a multi-step pipeline for converting comic pages to accessible literary narratives, including detection, association, OCR, panel subtitle generation, character localization, and text generation.
From Prompt to Progression: Taming Video Diffusion Models for Seamless Attribute Transition
Ling Lo (National Yang Ming Chiao Tung University), Ming-Hsuan Yang (UC Merced)
GenerationData SynthesisPrompt EngineeringDiffusion modelVideoBenchmark
🎯 What it does: This study proposes a training-independent framework that achieves smooth transitions of video attributes by incorporating frame-level guidance and neutral prompt anchoring into the denoising process of diffusion models.
From Reflection to Perfection: Scaling Inference-Time Optimization for Text-to-Image Diffusion Models via Reflection Tuning
Le Zhuo (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
GenerationOptimizationTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageText
🎯 What it does: The ReflectionFlow framework is proposed, which iteratively self-improves during diffusion model inference through three dimensions: noise, prompts, and reflection, and constructs a dataset of 1 million reflection triplets, GenRef, for reflection tuning.
From Reusing to Forecasting: Accelerating Diffusion Models with TaylorSeers
Jiacheng Liu (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
GenerationComputational EfficiencyTransformerDiffusion modelVideo
🎯 What it does: A cache-prediction paradigm based on Taylor expansion (TaylorSeer) is proposed to accelerate the inference of Diffusion Transformers.
From Sharp to Blur: Unsupervised Domain Adaptation for 2D Human Pose Estimation Under Extreme Motion Blur Using Event Cameras
Youngho Kim (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
Pose EstimationDomain AdaptationOptical FlowImageMultimodality
🎯 What it does: Using event cameras as a bridge, the annotations from the clear image domain are transferred to the motion-blurred image domain, achieving unsupervised domain adaptation for human pose estimation.
From Trial to Triumph: Advancing Long Video Understanding via Visual Context Sample Scaling and Self-reward Alignment
Yucheng Suo (Zhejiang University), Yi Yang (Zhejiang University)
RecognitionOptimizationTransformerLarge Language ModelReinforcement LearningVideoMultimodality
🎯 What it does: By randomly sampling key frames multiple times from long videos and generating multiple predictions, and then selecting the best answer using a self-reward mechanism (frequency, marginal confidence, context voting), the effectiveness of multimodal large language models in long video understanding is improved.
FROSS: Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images
Hao-Yu Hou (National Tsing Hua University), Yasutomo Kawanishi (RIKEN)
RecognitionObject DetectionGenerationComputational EfficiencySimultaneous Localization and MappingImagePoint CloudGraph
🎯 What it does: A framework for online, real-time 3D Semantic Scene Graph (SSG) generation called FROSS is proposed, which directly elevates 2D scene graphs to 3D and achieves incremental construction through Gaussian distribution fusion.
FullDiT: Video Generative Foundation Models with Multimodal Control via Full Attention
Xuan Ju (Kuaishou Technology), Qiang Xu (Chinese University of Hong Kong)
GenerationData SynthesisTransformerVideoTextMultimodality
🎯 What it does: This paper proposes FullDiT, a unified multi-task video generation foundation model that can integrate multimodal control conditions such as text, camera, identity, and depth within the same sequence representation.
Function-centric Bayesian Network for Zero-Shot Object Goal Navigation
Sixian Zhang (Institute of Computing Technology, Chinese Academy of Sciences), Shuqiang Jiang (Institute of Computing Technology, Chinese Academy of Sciences)
Object DetectionRobotic IntelligenceLarge Language ModelPrompt EngineeringPoint CloudChain-of-Thought
🎯 What it does: This paper proposes a Function-Centered Bayesian Network (FBN) that infers object-function and scene-function relationships through LLM, generating target probability graphs using Bayesian inference to guide zero-shot target navigation.
Fuse Before Transfer: Knowledge Fusion for Heterogeneous Distillation
Guopeng Li (Wuhan University), Gui-Song Xia (Wuhan University)
Knowledge DistillationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes a knowledge distillation framework called Fusion Before Transfer (FBT), which narrows the feature gap between cross-architecture models by mixing CNN and MSA/MLP modules.
Fusion Meets Diverse Conditions: A High-diversity Benchmark and Baseline for UAV-based Multimodal Object Detection with Condition Cues
Chen Chen (National University of Defense Technology), Ping Zhong (National University of Defense Technology)
Object DetectionPrompt EngineeringImageMultimodalityBenchmark
🎯 What it does: A highly diverse UAV RGB-IR multimodal object detection dataset, ATR-UMOD, is proposed, and a dynamic fusion method based on conditional prompts, PCDF, is developed to enhance detection performance under varying imaging conditions.
FusionPhys: A Flexible Framework for Fusing Complementary Sensing Modalities in Remote Physiological Measurement
Chenhang Ying (Zhejiang University), Xiaobai Li (Zhejiang University)
RecognitionData-Centric LearningConvolutional Neural NetworkVideoMultimodality
🎯 What it does: The FusionPhys framework is proposed, which adaptively fuses the temporal signals from three types of sensors: visible light, near-infrared, and radar, through a variable temporal modulation matrix, and enhances accuracy by decomposing single-modal videos into multiple sub-modalities using sub-modality decomposition.
Future-Aware Interaction Network For Motion Forecasting
Shijie Li (A*STAR), Xulei Yang (A*STAR)
Autonomous DrivingOptimizationTransformerTime Series
🎯 What it does: This paper proposes the Future-Aware Interaction Network (FINet), which models potential future trajectories in advance and integrates them into scene encoding to achieve joint optimization of historical and future states in motion prediction.
FuXi-RTM: A Physics-Guided Prediction Framework with Radiative Transfer Modeling
Qiusheng Huang (Fudan University), Hao Li (Fudan University)
Recurrent Neural NetworkTime SeriesPhysics Related
🎯 What it does: This paper proposes FuXi-RTM, a hybrid physical prediction framework that integrates deep learning with radiative transfer models to enhance the accuracy and physical consistency of global weather forecasts.
Fuzzy Contrastive Decoding to Alleviate Object Hallucination in Large Vision-Language Models
Jieun Kim (Yonsei University), Sung-Bae Cho (Yonsei University)
Object DetectionVision Language ModelContrastive LearningMultimodality
🎯 What it does: A Fuzzy Contrastive Decoding (FuzzyCD) decoding strategy based on fuzzy reasoning is proposed to alleviate the object hallucination problem in large visual language models.
FVGen: Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation
Wenbin Teng (University of Southern California), Yajie Zhao (University of Southern California)
GenerationData SynthesisComputational EfficiencyKnowledge DistillationDiffusion modelGenerative Adversarial NetworkVideo
🎯 What it does: The FVGen framework is proposed, which can quickly generate new views under sparse perspectives by distilling a multi-step video diffusion model, significantly reducing sampling time.
FW-Merging: Scaling Model Merging with Frank-Wolfe Optimization
Hao Mark Chen (Imperial College London), Hongxiang Fan (Imperial College London)
OptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: A model merging method based on Frank-Wolfe optimization (FW-Merging) is proposed, which iteratively selects the most relevant models and merges them using linear approximation to achieve efficient and scalable multi-task learning among massive black-box models.
G-DexGrasp: Generalizable Dexterous Grasping Synthesis Via Part-Aware Prior Retrieval and Prior-Assisted Generation
Juntao Jian (Shenzhen University), Ruizhen Hu (Shenzhen University)
GenerationRetrievalOptimizationRobotic IntelligenceTransformerLarge Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: A retrieval-enhanced generation framework G-DexGrasp is proposed to generate high-quality, multi-finger grasping hand poses under unseen object categories and natural language task instructions.
G2D: Boosting Multimodal Learning with Gradient-Guided Distillation
Mohammed Rakib (Oklahoma State University), Arunkumar Bagavathi (Oklahoma State University)
Knowledge DistillationMultimodality
🎯 What it does: The Gradient-Guided Distillation (G2D) framework is proposed, which addresses the modality imbalance problem in multimodal learning by combining knowledge distillation and sequential modality prioritization.
G2PDiffusion: Cross-Species Genotype-to-Phenotype Prediction via Evolutionary Diffusion
Mengdi Liu (Institute of Computing Technology, Chinese Academy of Sciences), Xilin Chen (Institute of Computing Technology, Chinese Academy of Sciences)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: A cross-species genome-to-phenotype image generation model G2PDiffusion is proposed, which predicts the morphological images of insects based on gene sequences, evolutionary information (MSA), and environmental factors.
G2SF: Geometry-Guided Score Fusion for Multimodal Industrial Anomaly Detection
Chengyu Tao (Hong Kong University of Science and Technology), Juan Du (Hong Kong University of Science and Technology)
Anomaly DetectionTransformerImageMultimodalityPoint Cloud
🎯 What it does: A geometry-guided multimodal industrial defect detection framework G2SF is proposed, which integrates anomaly scores from 3D point clouds and RGB images.
Gain-MLP: Improving HDR Gain Map Encoding via a Lightweight MLP
Trevor D. Canham (York University), Michael S. Brown (Rochester Institute of Technology)
RestorationCompressionImage
🎯 What it does: A lightweight MLP-based HDR gain map encoding method is proposed, using SDR images as input and training a 10 KB model to achieve HDR reconstruction.
Gait-X: Exploring X modality for Generalized Gait Recognition
Zengbin Wang (Beijing University of Posts and Telecommunications), Man Zhang (Beijing University of Posts and Telecommunications)
RecognitionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a gait recognition method based on frequency decomposition called Gait-X, which constructs a new X gait modality and designs a frequency domain training strategy to enhance both cross-domain and same-domain recognition performance.
GameFactory: Creating New Games with Generative Interactive Videos
Jiwen Yu, Xihui Liu
GenerationDomain AdaptationTransformerDiffusion modelVideo
🎯 What it does: Proposes the GameFactory framework, which utilizes pre-trained video diffusion models and a small number of game videos to achieve interactive game generation.
GAP: Gaussianize Any Point Clouds with Text Guidance
Weiqi Zhang (Tsinghua University), Yu-Shen Liu (Tsinghua University)
GenerationData SynthesisDiffusion modelGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes a framework named GAP, which can automatically generate high-quality, visually impressive 3D Gaussian point clouds from uncolored raw point clouds guided by text prompts.
GaRe: Relightable 3D Gaussian Splatting for Outdoor Scenes from Unconstrained Photo Collections
Haiyang Bai (Nanjing University), Lijun Chen (Nanjing University)
GenerationData SynthesisGaussian SplattingImage
🎯 What it does: A framework called GaRe is constructed based on 3D Gaussian Splatting to achieve real-time relighting from unconstrained outdoor photo collections. It utilizes an intrinsic image decomposition method to separate lighting into sunlight, sky radiation, and indirect illumination, achieving precise separation through visibility extraction, region supervision, and structural consistency loss. It also combines ray tracing with visibility baking to synthesize shadows in real-time.
GARF: Learning Generalizable 3D Reassembly for Real-World Fractures
Sihang Li (New York University), Jing Zhang (New York University)
SegmentationDomain AdaptationTransformerPoint Cloud
🎯 What it does: Proposes the GARF framework for generalizable 3D reconstruction of real-world fragments and creates the FRACTURA dataset containing ceramic, bone, eggshell, and stone tool fragments.
GAS: Generative Avatar Synthesis from a Single Image
Yixing Lu (Carnegie Mellon University), Fernando De la Torre (Carnegie Mellon University)
GenerationData SynthesisPose EstimationDiffusion modelNeural Radiance FieldAuto EncoderImageVideo
🎯 What it does: A unified framework for generating high-quality, view-consistent, and temporally coherent avatars from a single image is proposed.
GaSLight: Gaussian Splats for Spatially-Varying Lighting in HDR
Christophe Bolduc (Universite Laval), Jean-François Lalonde (Universite Laval)
GenerationData SynthesisDiffusion modelGaussian SplattingImage
🎯 What it does: Generate spatially varying HDR lighting from conventional photographic images, using HDR Gaussian Splats as light source representations that can be directly utilized in 3D renderers;
Gaussian Splatting with Discretized SDF for Relightable Assets
Zuo-Liang Zhu (Nankai University), Beibei Wang (Nanjing University)
Gaussian SplattingPoint Cloud
🎯 What it does: The paper proposes a method for reconstructing relightable assets using discretized SDF within a 3D Gaussian scattering framework.
Gaussian Variation Field Diffusion for High-fidelity Video-to-4D Synthesis
Bowen Zhang (University of Science and Technology of China), Baining Guo (Microsoft Research Asia)
GenerationData SynthesisTransformerDiffusion modelGaussian SplattingVideoMesh
🎯 What it does: This paper proposes a framework based on diffusion models that can generate high-quality, spatiotemporally consistent 4D animations (dynamic 3D content) from a single natural video.
Gaussian-based World Model: Gaussian Priors for Voxel-Based Occupancy Prediction and Future Motion Prediction
Tuo Feng (Zhejiang University), Yi Yang (Zhejiang University)
Autonomous DrivingOptimizationTransformerGaussian SplattingWorld ModelMultimodalityPoint Cloud
🎯 What it does: A Gaussian Prior-based World Model (GWM) is proposed, achieving an end-to-end unified framework for 4D occupancy prediction and future motion prediction from raw multimodal sensor inputs (cameras and LiDAR).
GaussianFlowOcc: Sparse and Weakly Supervised Occupancy Estimation using Gaussian Splatting and Temporal Flow
Simon Boeder (Robert Bosch GmbH), Benjamin Risse (University of Munster)
SegmentationAutonomous DrivingTransformerGaussian SplattingOptical FlowImageMultimodality
🎯 What it does: This paper proposes GaussianFlowOcc, which uses a sparse 3D Gaussian distribution to represent scenes and performs weakly supervised occupancy estimation through Temporal Gaussian Splatting and Temporal Flow.
GaussianOcc: Fully Self-supervised and Efficient 3D Occupancy Estimation with Gaussian Splatting
Wanshui Gan (University of Tokyo), Naoto Yokoya (University of Tokyo)
SegmentationAutonomous DrivingComputational EfficiencyTransformerGaussian SplattingPoint Cloud
🎯 What it does: Proposes the GaussianOcc method, which utilizes Gaussian splatting for unsupervised 3D occupancy estimation.
GaussianProperty: Integrating Physical Properties to 3D Gaussians with LMMs
Xinli Xu (Hong Kong University of Science and Technology), Ying-Cong Chen (Hong Kong University of Science and Technology)
SegmentationOptimizationRobotic IntelligenceTransformerLarge Language ModelGaussian SplattingImageMultimodalityPhysics Related
🎯 What it does: This paper proposes a training-free framework—GaussianProperty, which utilizes Segment Anything (SAM) for multi-level segmentation of objects. Subsequently, it employs GPT-4V (vision) to perform global-local integration of physical property inference (such as density, elastic modulus, friction coefficient, etc.) for each segmented fragment. The obtained 2D properties are then mapped onto the 3D Gaussian cloud generated from multi-view reconstruction using a voting strategy, achieving seamless assignment of physical properties to 3D models, and directly applying them to physics-based simulations (MPM, PhysGaussian) and robotic grasping force prediction.
GaussianReg: Rapid 2D/3D Registration for Emergency Surgery via Explicit 3D Modeling with Gaussian Primitives
Weihao Yu (Chinese University of Hong Kong), Yixuan Yuan (Chinese University of Hong Kong)
Pose EstimationComputational EfficiencyGaussian SplattingImageMultimodalityComputed Tomography
🎯 What it does: This paper presents GaussianReg, a fast 2D/3D registration framework for emergency surgery that can achieve high-precision registration in a matter of minutes.
GaussianSpeech: Audio-Driven Personalized 3D Gaussian Avatars
Shivangi Aneja (Technical University of Munich), Matthias Nießner (Technical University of Munich)
GenerationTransformerGaussian SplattingVideoAudio
🎯 What it does: Combining audio-driven 3D Gaussian splatting, we propose GaussianSpeech, which can generate high-fidelity, personalized 3D head animations from any viewpoint.
GaussianUpdate: Continual 3D Gaussian Splatting Update for Changing Environments
Lin Zeng (Zhejiang University), Zhaopeng Cui (Zhejiang University)
SegmentationGenerationOptimizationGaussian SplattingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: The GaussianUpdate method is proposed, which utilizes 3D Gaussian splatting combined with continual learning to update the scene in real-time and render new views at different time points.
GaussianVideo: Efficient Video Representation via Hierarchical Gaussian Splatting
Andrew Bond (Koc University), Aykut Erdem (Koc University)
RestorationCompressionRepresentation LearningGaussian SplattingVideoOrdinary Differential Equation
🎯 What it does: This paper presents GaussianVideo, a video representation framework based on hierarchical 3D Gaussian splatting and neural ODE, achieving unsupervised efficient video reconstruction.
GausSim: Foreseeing Reality by Gaussian Simulator for Elastic Objects
Yidi Shao (Nanyang Technological University), Bo Dai (University of Hong Kong)
GenerationComputational EfficiencyGaussian SplattingVideoPhysics Related
🎯 What it does: This paper presents GausSim, a physics simulator based on neural networks for dynamic simulation of elastic objects represented by Gaussian Splatting.
GaussRender: Learning 3D Occupancy with Gaussian Rendering
Loick Chambon, Matthieu Cord (Sorbonne University)
Autonomous DrivingGaussian SplattingPoint Cloud
🎯 What it does: The GaussRender module is proposed, which achieves 2D projection consistency supervision for 3D occupancy prediction through differentiable Gaussian light scattering rendering, significantly improving geometric consistency and surface localization.
GauUpdate: New Object Insertion in 3D Gaussian Fields with Consistent Global Illumination
Chengwei Ren (Tsinghua University), Yuan Liu (Hong Kong University of Science and Technology)
GenerationData SynthesisNeural Radiance FieldImagePoint Cloud
🎯 What it does: Proposes the GauUpdate framework to insert new objects into an already constructed large-scale 3D Gaussian field, achieving global illumination consistency through inverse rendering, thus avoiding the need to reconstruct the entire scene.
Gaze-Language Alignment for Zero-Shot Prediction of Visual Search Targets from Human Gaze Scanpaths
Sounak Mondal (Stony Brook University), Tanya R. Jonker (Meta Reality Labs Research)
RecognitionObject DetectionTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodality
🎯 What it does: This paper proposes a visual-language model named GLAM, which predicts target categories from human gaze trajectories in visual search tasks and supports zero-shot prediction.
GazeGaussian: High-Fidelity Gaze Redirection with 3D Gaussian Splatting
Xiaobao Wei (Institute of Software, Chinese Academy of Sciences), Feng Tian (Institute of Software, Chinese Academy of Sciences)
GenerationData SynthesisGaussian SplattingImage
🎯 What it does: A high-fidelity gaze redirection method called GazeGaussian based on dual-stream 3D Gaussian splatting is proposed, achieving precise separation and control of the face and eyes.
GCAV: A Global Concept Activation Vector Framework for Cross-Layer Consistency in Interpretability
Zhenghao He (University of Virginia), Aidong Zhang (University of Virginia)
Explainability and InterpretabilityConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Proposes the GCAV framework to unify cross-layer concept activation vectors and use them for interpretation.
GCRayDiffusion: Pose-Free Surface Reconstruction via Geometric Consistent Ray Diffusion
Li-Heng Chen (Beijing Normal University), Hua Huang (Beijing Normal University)
RestorationSegmentationPose EstimationDepth EstimationDiffusion modelPoint CloudMesh
🎯 What it does: This paper proposes a posture-free 3D surface reconstruction method based on Geometric Consistent Ray Diffusion (GCRayDiffusion), which can simultaneously achieve camera pose estimation and fine surface reconstruction under sparse viewpoint images.
GDKVM: Echocardiography Video Segmentation via Spatiotemporal Key-Value Memory with Gated Delta Rule
Rui Wang (Shenzhen University), Jing Qin (Hong Kong Polytechnic University)
SegmentationConvolutional Neural NetworkVideoBiomedical DataUltrasound
🎯 What it does: A linear key-value memory network GDKVM is proposed for real-time segmentation of cardiac ultrasound videos.
GECKO: Gigapixel Vision-Concept Contrastive Pretraining in Histopathology
Saarthak Kapse (Stony Brook University), Prateek Prasanna (Stony Brook University)
ClassificationExplainability and InterpretabilityRepresentation LearningLarge Language ModelVision Language ModelContrastive LearningImageBiomedical Data
🎯 What it does: Using an unsupervised contrastive learning method, a dual-branch multi-instance learning (MIL) model is pre-trained to align pathological WSIs with text-based conceptual priors, thereby obtaining interpretable WSI-level embeddings.
GECO: Geometrically Consistent Embedding with Lightspeed Inference
Regine Hartwig (Technical University of Munich), Daniel Cremers (Technical University of Munich)
RecognitionPose EstimationTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: Learn geometrically consistent visual features through optimal transport (OT) soft assignment, addressing the challenges of symmetric part differentiation while considering occlusion issues.
GEMeX: A Large-Scale, Groundable, and Explainable Medical VQA Benchmark for Chest X-ray Diagnosis
Bo Liu (Hong Kong Polytechnic University), Huazhu Fu (Institute of High Performance Computing)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBiomedical DataBenchmark
🎯 What it does: A large, interpretable, and locatable chest X-ray visual question answering benchmark, GEMeX, has been constructed, providing various types of questions.
Geminio: Language-Guided Gradient Inversion Attacks in Federated Learning
Junjie Shan (University of Hong Kong), Ka-Ho Chow (University of Hong Kong)
Federated LearningAdversarial AttackTransformerVision Language ModelImage
🎯 What it does: This paper proposes Geminio, which utilizes a pre-trained visual-language model (VLM) to perform semantic guidance for gradient inversion attacks in federated learning, achieving targeted recovery of specified samples.
GenDoP: Auto-regressive Camera Trajectory Generation as a Director of Photography
Mengchen Zhang (Zhejiang University), Dahua Lin
GenerationData SynthesisTransformerVision Language ModelImageVideoMultimodality
🎯 What it does: A GenDoP model based on autoregressive Transformer is proposed to automatically generate artistic and freely moving camera trajectories conditioned on text or RGBD;
General Compression Framework for Efficient Transformer Object Tracking
Lingyi Hong (Fudan University), Wenqiang Zhang (Fudan University)
Object TrackingCompressionComputational EfficiencyKnowledge DistillationTransformerVideo
🎯 What it does: This paper presents CompressTracker, a universal model compression framework designed to compress large-scale Transformer trackers into efficient student models while retaining nearly the same tracking performance.
Generalizable Non-Line-of-Sight Imaging with Learnable Physical Priors
Shida Sun (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
RestorationDepth EstimationConvolutional Neural NetworkImage
🎯 What it does: A learning-based non-line-of-sight imaging method is proposed, which includes two main modules: learnable path compensation (LPC) and adaptive phase field (APF). This method can achieve high-quality 3D reconstruction under low signal-to-noise ratio (SNR) conditions and has good generalization ability for real data.
Generalizable Object Re-Identification via Visual In-Context Prompting
Zhizhong Huang (Michigan State University), Xiaoming Liu (Michigan State University)
RecognitionRetrievalTransformerLarge Language ModelPrompt EngineeringImage
🎯 What it does: This paper proposes a Visual Context Prompting Framework (VICP) that utilizes large language models and visual foundation models to achieve object re-identification for unseen categories without the need for additional parameter fine-tuning.
Generalization-Preserved Learning: Closing the Backdoor to Catastrophic Forgetting in Continual Deepfake Detection
Xueyi Zhang (National University of Defense Technology), Yanming Guo (Harbin Institute of Technology)
ClassificationAnomaly DetectionConvolutional Neural NetworkContrastive LearningVideo
🎯 What it does: A continuous deepfake detection framework based on Generalization-Preserved Learning (GPL) is proposed, aiming to enhance both the stability and plasticity of the model while avoiding catastrophic forgetting.
Generalized and Efficient 2D Gaussian Splatting for Arbitrary-scale Super-Resolution
Du Chen (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
RestorationSuper ResolutionConvolutional Neural NetworkGaussian SplattingImage
🎯 What it does: A Gaussian Splatting-based arbitrary scale super-resolution model GSASR is proposed, which can convert low-resolution images into continuous Gaussian representations and generate high-resolution images at arbitrary magnification factors through differentiable GPU/CUDA rasterization.
Generalized Deep Multi-view Clustering via Causal Learning with Partially Aligned Cross-view Correspondence
Xihong Yang (National University of Defense Technology), Kunlun He (Chinese PLA General Hospital)
Auto EncoderContrastive LearningMultimodality
🎯 What it does: A deep clustering method for partially aligned multi-view data, CauMVC, is proposed, utilizing causal learning to achieve generalization from fully aligned to partially aligned.
Generalized Few-Shot Point Cloud Segmentation via LLM-Assisted Hyper-Relation Matching
Zhaoyang Li (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
SegmentationKnowledge DistillationGraph Neural NetworkLarge Language ModelPoint Cloud
🎯 What it does: A large language model (LLM) assisted hyper-relation matching framework (LARM) is proposed for general few-shot point cloud semantic segmentation.
Generalized Tensor-based Parameter-Efficient Fine-Tuning via Lie Group Transformations
Chongjie Si (Shanghai Jiao Tong University), Wei Shen (Shanghai Jiao Tong University)
ClassificationObject DetectionGenerationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper presents LieRA—a parameter-efficient fine-tuning framework based on Lie group theory, which extends traditional matrix-based PEFT methods (such as LoRA) to high-dimensional tensors (e.g., convolution kernels) while preserving their spatial structure.
Generate, Refine, and Encode: Leveraging Synthesized Novel Samples for On-the-Fly Fine-Grained Category Discovery
Xiao Liu (Hefei University of Technology), Zhun Zhong (University of Trento)
ClassificationGenerationData SynthesisDiffusion modelContrastive LearningImage
🎯 What it does: This paper studies the task of Online Category Discovery (OCD) and proposes the DiffGRE framework, which generates virtual category samples by reorganizing attributes in latent space through a diffusion model, enhancing the generalization ability of the OCD model with these samples.