These 833 ICCV 2025 papers come with a code repository. Each shows an AI one-line summary below β get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every ICCV 2025 paper, free trial on arXivSub.
3D Gaussian Map with Open-Set Semantic Grouping for Vision-Language Navigation
Jianzhe Gao (Zhejiang University), Wenguan Wang (Zhejiang University)
π― What it does: This paper proposes a scene map (Egocentric Scene Map) constructed based on sparse differentiable 3D Gaussian atoms, and assigns open semantic labels to each atom through Open-Set Semantic Grouping, ultimately enhancing the decision-making ability of visual language navigation (VLN) agents using multi-level action prediction (Scene, View, Instance).
π― What it does: A physical adversarial attack framework based on 3D Gaussian Splatting (PGA) has been developed, capable of generating multi-view robust adversarial camouflage for arbitrary targets in both digital and real environments.
Yanyan Li (Hangzhou Dianzi University), Federico Tombari (Google)
CodeObject TrackingPose EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingOptical FlowPoint Cloud
π― What it does: A 4D Gaussian Splatting SLAM system is proposed, capable of simultaneously estimating camera trajectories and constructing dynamic Gaussian radiance fields in dynamic scenes.
6DOPE-GS: Online 6D Object Pose Estimation using Gaussian Splatting
Yufeng Jin (Technische Universitat Darmstadt), Georgia Chalvatzaki (Technische Universitat Darmstadt)
CodePose EstimationGaussian SplattingSimultaneous Localization and MappingPoint Cloud
π― What it does: This paper proposes an online 6D pose estimation and reconstruction framework called 6DOPE-GS based on 2D Gaussian distribution projection rendering.
A Framework for Double-Blind Federated Adaptation of Foundation Models
Nurbek Tastan (Mohamed bin Zayed University of Artificial Intelligence), Karthik Nandakumar (Michigan State University)
CodeFederated LearningSafty and PrivacyKnowledge DistillationTransformerImage
π― What it does: This paper proposes the BlindFed framework, which enables multiple data owners to collaboratively adapt large foundational models in a double-blind federated manner without sharing models or data with the learning service provider.
A Hidden Stumbling Block in Generalized Category Discovery: Distracted Attention
Qiyu Xu (Yunnan Normal University), Yonghang Tai (Yunnan Normal University)
CodeClassificationRecognitionTransformerImage
π― What it does: Proposes the Attention Focusing (AF) module, which utilizes two components, TIME and TAP, to adaptively prune irrelevant tokens in the Vision Transformer, thereby reducing distracting attention and enhancing feature quality in Generalized Category Discovery (GCD).
A Hyperdimensional One Place Signature to Represent Them All: Stackable Descriptors For Visual Place Recognition
Connor Malone (Queensland University of Technology), Michael Milford (Queensland University of Technology)
CodeRecognitionAutonomous DrivingSimultaneous Localization and MappingImage
π― What it does: This paper proposes a Hyperdimensional One Place Signatures (HOPS) framework based on high-dimensional computation, which fuses VPR descriptors of the same location under different conditions to enhance the robustness of visual localization.
CodeRecognitionSegmentationRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: This paper develops the TokenIT dataset, the TokenFD text-image foundational model, and the TokenVL multimodal large language model for fine semantic understanding and reasoning of text images such as documents, tables, and charts.
π― What it does: A unified BRIDGE framework is proposed to bridge complete multi-view clustering (DMVC) and incomplete multi-view clustering (DIMVC), taking into account different multi-view interaction strategies and addressing the issue of non-IID missing patterns.
π― What it does: Improved 3D Gaussian projection rendering, proposing the use of complete 3D Gaussian evaluation throughout the entire rendering process to eliminate visual defects such as aliasing, projection distortion, jitter, and popping.
π― What it does: This paper proposes a post-training OOD detection method named ActSub, which utilizes singular value decomposition of the classification head weight matrix to decompose network activations into deterministic and non-deterministic subspaces, and calculates similarity and energy scores separately in both, ultimately merging them to obtain an OOB discrimination score.
Active Learning Meets Foundation Models: Fast Remote Sensing Data Annotation for Object Detection
Marvin Burges (TU Wien), Dalton Lunga (Oak Ridge National Laboratory)
CodeObject DetectionImage
π― What it does: A real-time active learning framework is proposed, utilizing the base model SAM to generate masks combined with detection boxes, achieving efficient labeling of semi-automated remote sensing image target detection datasets.
π― What it does: This paper proposes a two-stage object navigation framework called APRR, which first uses rule-guided reinforcement learning to efficiently explore unknown environments, and then employs active perception reinforcement learning to refine the precise docking position at the target.
AdaDrive: Self-Adaptive Slow-Fast System for Language-Grounded Autonomous Driving
Ruifei Zhang (Chinese University of Hong Kong), Guanbin Li (Sun Yat-sen University)
CodeAutonomous DrivingTransformerLarge Language ModelReinforcement LearningText
π― What it does: AdaDrive achieves on-demand activation and dynamic fusion of LLM in autonomous driving through an adaptive slow-fast architecture, enhancing decision quality.
π― What it does: To address the cross-domain few-shot segmentation problem, we propose an adaptive structural adjustment of the pre-trained FSS model during inference using a small number of labeled support samples, enabling cross-domain transfer without retraining on the source domain.
π― What it does: This paper proposes the DUO framework, which utilizes Test-Time Adaptation (TTA) technology to simultaneously minimize semantic uncertainty and geometric uncertainty in monocular 3D object detection, thereby enhancing detection performance in domain shift environments.
Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections
Youwei Zhou (Jiangnan University), Josef Kittler (University of Surrey)
CodeRecognitionGraph Neural NetworkVideo
π― What it does: An adaptive hypergraph convolutional network (Hyper-GCN) is proposed, which constructs variable multi-vertex hypergraphs and introduces virtual connections to achieve efficient feature aggregation for skeleton sequences, thereby completing human action recognition based on skeletons.
π― What it does: A semi-supervised medical image segmentation framework called ALHVR based on high-value region adaptive learning is proposed. It utilizes a dual-branch network to predict differences and classify pixels into three categories: reliable stable, reliable unstable, and unreliable stable regions, and designs an adaptive learning strategy specifically for high-value regions.
Xinwei Long (Tsinghua University), Bowen Zhou (Tsinghua University)
CodeTransformerLarge Language ModelReinforcement LearningVideoMultimodalityBenchmark
π― What it does: Designed and constructed the AdsQA advertising video question-answering benchmark, and proposed the ReAd-R model based on reinforcement learning to evaluate the implicit reasoning ability of LLMs in advertising videos.
Advancing Textual Prompt Learning with Anchored Attributes
Zheng Li (Nankai University), Jian Yang (Nankai University)
CodeClassificationDomain AdaptationTransformerPrompt EngineeringVision Language ModelImageMultimodality
π― What it does: A text prompting method based on general attribute anchoring (ATPrompt) is designed for visual-language models, enhancing the alignment capability between images and unknown categories by embedding attribute words into soft prompts.
CodeGenerationPose EstimationAdversarial AttackTransformerVision Language ModelDiffusion modelImageTextBenchmark
π― What it does: The AdvDreamer framework is proposed, which can generate adversarial 3D transformation (Adv-3DT) samples that can be reproduced in the real physical world using only single-view images, and thus evaluate the robustness of visual language models (VLM) under dynamic 3D changes.
π― What it does: Proposes the LEAwareSGD optimizer, which guides training close to the edge of chaos through the Lyapunov exponent, achieving data augmentation for single-domain generalization.
π― What it does: A method for adversarial training aimed at probabilistic robustness, called AT-PR, is proposed to enhance the model's probability (PR) of making incorrect classifications within a perturbation ball.
π― What it does: This paper proposes the Aerial Visual Grounding (AerialVG) task, constructs the first aerial visual grounding dataset containing 5K high-resolution aerial images, 50K natural language descriptions, and 103K objects, and presents a dedicated model to achieve this task.
AFUNet: Cross-Iterative Alignment-Fusion Synergy for HDR Reconstruction via Deep Unfolding Paradigm
Xinyue Li (Tongji University), Wenhan Yang (Pengcheng Laboratory)
CodeRestorationTransformerImage
π― What it does: This paper proposes a cross-iteration alignment-fusion deep unfolding network (AFUNet) for reconstructing high dynamic range (HDR) images from multi-exposure low dynamic range images.
AGO: Adaptive Grounding for Open World 3D Occupancy Prediction
Peizheng Li (Mercedes-Benz AG), Andreas Zell (University of TΓΌbingen)
CodeRecognitionObject DetectionSegmentationAutonomous DrivingTransformerVision Language ModelContrastive LearningImagePoint Cloud
π― What it does: Proposes the AGO framework, which is based on VLM knowledge self-supervised learning for 3D semantic occupancy prediction, and achieves recognition of known and unknown objects in open-world scenarios.
π― What it does: An improved GLOM model is proposed, incorporating a contrastive consistency enhancer and a diversity focus head to enhance the formation of embedding islands and the computational efficiency of the model.
π― What it does: A post-training quantization framework AHCPTQ is proposed for the Segment Anything Model (SAM), addressing the issues of activation distribution skew and excessive differences between channels in SAM quantization.
π― What it does: Proposes AIComposer, which achieves cross-domain image synthesis using linear/non-linear fusion of CLIP features without the need for text prompts;
AIGI-Holmes: Towards Explainable and Generalizable AI-Generated Image Detection via Multimodal Large Language Models
Ziyin Zhou (Xiamen University), Rongrong Ji (Xiamen University)
CodeObject DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningImageMultimodality
π― What it does: An explainable and generalizable AI-generated image detection method called AIGI-Holmes based on a multimodal large language model is proposed, along with the construction of the Holmes-Set dataset, which provides human-verifiable explanations.
AIM: Adaptive Inference of Multi-Modal LLMs via Token Merging and Pruning
Yiwu Zhong (Chinese University of Hong Kong), Liwei Wang (Chinese University of Hong Kong)
CodeComputational EfficiencyTransformerLarge Language ModelVision Language ModelImageVideoMultimodality
π― What it does: An adaptive reasoning method that is training-independent is proposed, which significantly reduces computational load by first merging visual tokens in a multimodal large language model and then performing token pruning at the LLM layer.
π― What it does: AIRA is proposed, a low-rank adaptation framework based on activation information, which improves the initialization, hierarchical rank allocation, and training process of LoRA to enhance the efficiency of fine-tuning large models.
Aligning Global Semantics and Local Textures in Generative Video Enhancement
Zhikai Chen (University of Science and Technology of China), Tao Mei (HiDream.ai Inc.)
CodeRestorationGenerationDiffusion modelVideo
π― What it does: The Generative Video Enhancement framework GenVE, based on diffusion models, utilizes high-quality image references to achieve global semantic and local texture alignment for low-quality videos, enhancing video details and visual quality.
Yogesh Kumar (Indian Institute of Technology Jodhpur), Anand Mishra (Microsoft)
CodeRetrievalTransformerVision Language ModelVideo
π― What it does: The paper proposes a dual-stage sequence alignment model called MATR based on Transformer, which can accurately locate semantically matching segments in the target video based on video queries.
Aligning Vision to Language: Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning
Junming Liu (Tongji University), Botian Shi (New York University)
CodeClassificationRetrievalCompressionTransformerLarge Language ModelMixture of ExpertsVision Language ModelImageTextMultimodality
π― What it does: Construct a multimodal knowledge graph (MMKG) without manual annotation, generating image descriptions through a vision-language model and integrating them with textual information, thereby enhancing the cross-modal reasoning ability of large language models (LLMs).
Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation
Shuchang Ye (University of Sydney), Jinman Kim (University of Sydney)
CodeSegmentationVision Language ModelImageBiomedical Data
π― What it does: ProLearn framework is proposed, achieving medical image segmentation through prototype-driven semantic approximation without the need for text input during the inference phase.
π― What it does: This paper proposes the Allowable Oscillation Quantization (AOQ) method, which encourages weight quantization threshold oscillation in the early training phase and suppresses oscillation in the later phase, thereby expanding the solution space for low-bit-width quantization and improving model accuracy.
AMDANet: Attention-Driven Multi-Perspective Discrepancy Alignment for RGB-Infrared Image Fusion and Segmentation
Haifeng Zhong (Jilin University), Yixing Gao (Chinese Academy of Sciences)
CodeSegmentationConvolutional Neural NetworkImage
π― What it does: Proposes AMDANet, an attention-driven multi-view disparity alignment network for RGB-infrared image fusion and semantic segmentation.
An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval
Jaeseok Byun (Seoul National University), Taesup Moon (Seoul National University)
CodeRetrievalVision Language ModelContrastive LearningImageText
π― What it does: This paper proposes a post-processing framework RTD to reduce the task gap of text encoders in zero-shot synthesized image retrieval tasks.
π― What it does: A reverse-based memory measurement method called InvMM is proposed, which quantifies the degree of memory retention of a single image by utilizing the sensitive noise distribution of diffusion models.
π― What it does: This paper proposes a training-free, zero-shot autoregressive image editing method called ISLock, which utilizes Anchor Token Matching (ATM) to implicitly lock the structure during the decoding process, enabling tasks such as object replacement, addition, removal, attribute modification, and style transfer.
Anti-Tamper Protection for Unauthorized Individual Image Generation
Zelin Li (University of Illinois Urbana-Champaign), Dong Wang (University of Illinois Urbana-Champaign)
CodeGenerationData SynthesisSafty and PrivacyAdversarial AttackConvolutional Neural NetworkDiffusion modelAuto EncoderImage
π― What it does: A tamper-resistant perturbation (ATP) mechanism is proposed, which can inject protective perturbations into images to resist personalized image generation, and can detect when perturbations are tampered with during purification attacks, thereby rejecting illegal generation requests.
Any-SSR: How Recursive Least Squares Works in Continual Learning of Large Language Model
Kai Tong (South China University of Technology), Huiping Zhuang (South China University of Technology)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: The Analytic Subspace Routing (Any-SSR) framework is proposed to achieve continuous learning of large language models without using historical data. It avoids catastrophic forgetting by freezing low-level features, training independent LoRA subspaces for each task, and using a recursive least squares algorithm to construct task routers.
π― What it does: A single-image camera calibration method named AnyCalib is proposed, supporting perspective, distortion, and editing (cropping/stretching) of images, and is independent of the camera model.
π― What it does: A framework called AR-1-to-3 is proposed for generating consistent 3D objects from single-view images, utilizing a diffusion model to progressively predict new views in order of distance and generate complete multi-view images.
AR-VRM: Imitating Human Motions for Visual Robot Manipulation with Analogical Reasoning
Dejie Yang (Peking University), Yang Liu (Peking University)
CodePose EstimationRobotic IntelligenceTransformerVision Language ModelVideoMultimodality
π― What it does: This paper proposes AR-VRM, which utilizes large-scale human hand keypoint video to learn explicit action knowledge and maps human hand actions to robot execution instructions through similar reasoning, thereby completing visual robotic manipulation tasks.
π― What it does: This paper proposes a real-time interactive head generation framework based on continuous autoregression (AR) called ARIG, designed to generate high-quality and natural head movements in two-person dialogues in real-time.
ART: Adaptive Relation Tuning for Generalized Relation Prediction
Gopika Sudhakaran (TU Darmstadt), Stefan Roth (TU Darmstadt)
CodeClassificationDomain AdaptationTransformerSupervised Fine-TuningVision Language ModelImageMultimodality
π― What it does: This paper proposes the Adaptive Relation Tuning (ART) framework for visual relationship detection (VRD), which fine-tunes visual language models (VLM) for relationship classification through instruction tuning and adaptive sampling, enhancing the model's reasoning and generalization capabilities on unseen relationships.
π― What it does: This paper presents ArtEditor, an end-to-end image editing framework based on Diffusion Transformer, capable of learning from only a few examples and performing instruction-driven edits for specific styles.
Arti-PG: A Toolbox for Procedurally Synthesizing Large-Scale and Diverse Articulated Objects with Rich Annotations
Jianhua Sun (Shanghai Jiao Tong University), Cewu Lu (Shanghai Jiao Tong University)
CodeData SynthesisPose EstimationPoint CloudMesh
π― What it does: This work proposes the Arti-PG toolbox, which can automatically synthesize thousands of diverse 3D articulated objects based on procedural structure descriptions and point correspondence techniques, providing rich annotations.
Ask and Remember: A Questions-Only Replay Strategy for Continual Visual Question Answering
Imad Eddine Marouf (Institut Polytechnique de Paris), Joost Van De Weijer
CodeSafty and PrivacyKnowledge DistillationTransformerImageText
π― What it does: Proposes a method for memory replay that only uses past task questions and achieves continual learning in visual question answering through attention consistency distillation.
Asynchronous Event Error-Minimizing Noise for Safeguarding Event Dataset
Ruofei Wang (Hong Kong Baptist University), Renjie Wan (Hong Kong Baptist University)
CodeAnomaly DetectionSafty and PrivacyAdversarial AttackConvolutional Neural NetworkTime Series
π― What it does: A method for generating Unlearnable Examples for Asynchronous Event Streams (UEVs) is proposed, which ensures that event datasets are not used without authorization by constructing a noise suppression model that minimizes errors.
ATCTrack: Aligning Target-Context Cues with Dynamic Target States for Robust Vision-Language Tracking
Xiaokun Feng (University of Chinese Academy of Sciences), Kaiqi Huang (University of Chinese Academy of Sciences)
CodeObject TrackingTransformerLarge Language ModelImageVideoTextMultimodality
π― What it does: Proposes the ATCTrack tracker, which enhances the robustness of visual-language tracking by utilizing dynamically aligned multimodal target-context information.
π― What it does: This paper proposes Att-Adapter, a pluggable module that enables continuous and multi-attribute fine control over pre-trained text-to-image diffusion models without the need for paired data.
π― What it does: This paper proposes a general neural plagiarism attack framework based on diffusion models, capable of copying copyrighted images without training or fine-tuning the model, and achieving forgery and ambiguity attacks in copyright protection systems by damaging or replacing watermarks (both visible and invisible).
Attention to Trajectory: Trajectory-Aware Open-Vocabulary Tracking
Yunhao Li (Institute of Software Chinese Academy of Sciences), Libo Zhang (Institute of Software Chinese Academy of Sciences)
CodeObject TrackingTransformerLarge Language ModelContrastive LearningVideo
π― What it does: This paper proposes TRACT, an open vocabulary multi-object tracker that utilizes Trajectory Consistency Reinforcement (TCR) and Trajectory Enhanced Classification (TraCLIP) for trajectory awareness.
AU-Blendshape for Fine-grained Stylized 3D Facial Expression Manipulation
Hao Li (Beihang University), Junjun Pan (Beihang University)
CodeGenerationTransformerAuto EncoderImageMesh
π― What it does: This paper constructs the AUBlendSet dataset based on AU-Blendshape and the AUBlendNet model, achieving fine-grained, stylized 3D facial expression manipulation for any identity.
Automated Red Teaming for Text-to-Image Models through Feedback-Guided Prompt Iteration with Vision-Language Models
Wei Xu (Wuhan University), Lina Wang (Wuhan University)
CodeGenerationAdversarial AttackTransformerPrompt EngineeringVision Language ModelImageTextMultimodalityChain-of-Thought
π― What it does: Designed and implemented an automatic red team framework FGPI based on a visual language model, used for iteratively optimizing aggressive prompts for text-to-image models, capable of inducing the generation of non-compliant images without leaking harmful text.
π― What it does: A self-regressive denoising score matching (ADSM) framework is proposed, which uses a noise conditional score transformer to estimate the Stein score of videos and integrates scene, motion, and appearance information to achieve video anomaly detection.
Auxiliary Prompt Tuning of Vision-Language Models for Few-Shot Out-of-Distribution Detection
Wenjun Miao (Beihang University), Xiao Bai (Beihang University)
CodeAnomaly DetectionTransformerPrompt EngineeringVision Language ModelImageMultimodality
π― What it does: Proposes the Auxiliary Prompt Tuning (APT) framework, which enhances CLIP-based few-shot OOD detection using pseudo OOD samples from external auxiliary data.
B-VLLM: A Vision Large Language Model with Balanced Spatio-Temporal Tokens
Zhuqiang Lu (University of Sydney), Kun Hu (Edith Cowan University)
CodeTransformerLarge Language ModelVision Language ModelVideoText
π― What it does: Proposes the B-VLLM framework, which utilizes text-conditioned adaptive frame selection, temporal frame merging, and spatial visual token sampling/merging to achieve dynamic balance of video spatiotemporal information and control the number of visual tokens.
π― What it does: A backdoor defense framework named ESTI is proposed, which utilizes data tampered by attackers along with a 'poisoning model' to partition the dataset. It then isolates the identified backdoor samples into a dedicated class using 'trap labels', ultimately training a model that is ineffective against backdoors while maintaining high accuracy.
π― What it does: This paper proposes a post-training debiasing method called D3, which aims to remove backdoors in pre-trained models by maximizing the distance between model weights and the original backdoored model while maintaining low loss on clean data.
Background Invariance Testing According to Semantic Proximity
Zukang Liao (University of Oxford), Min Chen (University of Oxford)
CodeClassificationRecognitionImage
π― What it does: A visualization-based framework for background invariance testing is proposed, using associated ontologies to extend keywords for background scene sampling to evaluate the background invariance of machine learning models.
Balancing Conservatism and Aggressiveness: Prototype-Affinity Hybrid Network for Few-Shot Segmentation
Tianyu Zou (Wuhan University of Technology), Yi Rong (Wuhan University of Technology)
CodeRecognitionSegmentationTransformerImage
π― What it does: This study focuses on few-shot semantic segmentation and proposes the Prototype-Affinity Hybrid Network (PAHNet), which improves foreground recognition and background suppression by combining conservative prototype learning with aggressive affinity learning, utilizing the PFE and ASC modules.
π― What it does: A unified image fusion framework TITA is proposed, which balances task-invariant interaction and task-specific adaptation, enabling the completion of various fusion tasks without the need for task identification.
π― What it does: A Bidirectional Aggregation Network (BANet) is proposed, which achieves high-quality stereo matching on mobile devices using only 2D convolutions, capable of preserving detail and edge information.
Benchmarking Burst Super-Resolution for Polarization Images: Noise Dataset and Analysis
Inseung Hwang (Korea Advanced Institute of Science and Technology), Min H. Kim (Korea Advanced Institute of Science and Technology)
CodeRestorationSuper ResolutionImageBenchmark
π― What it does: This paper presents two publicly available datasets, PolarNS and PolarBurstSR, specifically designed for noise statistics and super-resolution of polarized images, and conducts theoretical modeling and empirical validation of polarized noise.
Benchmarking Multimodal Large Language Models Against Image Corruptions
Xinkuan Qiu (Institute of Information Engineering, Chinese Academy of Sciences), Shiguang Shan (Institute of Computing Technology, Chinese Academy of Sciences)
CodeTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
π― What it does: A robustness evaluation benchmark MLLM-IC for multimodal large language models under image distortion conditions is proposed.
π― What it does: This paper proposes DiffBrush, a diffusion model for generating handwritten text lines; it achieves style imitation of the overall text line and control over character readability through content-decoupled style learning and a multi-scale content discriminator.
π― What it does: The LiMA framework is proposed, which achieves more robust representation learning for LiDAR through cross-view aggregation, long-term feature propagation, and cross-sequence memory alignment for knowledge distillation from image to LiDAR.
Beyond Perspective: Neural 360-Degree Video Compression
Andy Regensky (Friedrich-Alexander-UniversitΓ€t Erlangen-NΓΌrnberg), Andre Kaup (Friedrich-Alexander-UniversitΓ€t Erlangen-NΓΌrnberg)
CodeCompressionOptical FlowVideo
π― What it does: A neural network compression framework for 360-degree video has been developed, combining large-scale datasets, spherical reprojection enhancement, and positional encoding to achieve bidirectional optimization for both 360-degree and traditional perspective videos.
π― What it does: This paper proposes the Raw Adaptation Module (RAM), a parallel ISP preprocessing module for end-to-end training on RAW images to enhance object detection performance.
π― What it does: This paper proposes an unsupervised hidden object detection framework called RISE, which generates pseudo-labels using a retrieval-incremental strategy to train the COD model.
Beyond the Destination: A Novel Benchmark for Exploration-Aware Embodied Question Answering
Kaixuan Jiang (Sun Yat-sen University), Liang Lin (Sun Yat-sen University)
CodeTransformerLarge Language ModelVision Language ModelTextMultimodalityBenchmark
π― What it does: A large-scale EQA benchmark dataset called EXPRESS-Bench has been constructed, and a Fine-EQA hybrid exploration model has been proposed.
Beyond Training: Dynamic Token Merging for Zero-Shot Video Understanding
Yiming Zhang (University of Science and Technology of China), Yining Sun (HFIPS)
CodeRecognitionCompressionComputational EfficiencyTransformerVision Language ModelVideo
π― What it does: A zero-shot video understanding framework called DYTO is proposed, which can efficiently and semantically encode videos without additional fine-tuning.
π― What it does: This paper proposes a text-driven pedestrian anomaly search task, constructs a large-scale Pedestrian Anomaly Behavior (PAB) visual-language benchmark, and introduces a Cross-Modal Pose-aware (CMP) framework.
π― What it does: A self-supervised AI-generated face detection method based on dual-layer optimization is proposed, which trains a visual encoder using EXIF tags and facial manipulation tasks without any manually synthesized samples.
π― What it does: Using normalizing flow to model the pixel feature distribution of the entire dataset, constructing distribution-based CAM and reliable feature sampling, thereby enhancing the robustness and accuracy of weakly supervised semantic segmentation.
Bidirectional Likelihood Estimation with Multi-Modal Large Language Models for Text-Video Retrieval
Dohwan Ko (Korea University), Hyunwoo J. Kim (KAIST)
CodeRetrievalTransformerLarge Language ModelSupervised Fine-TuningVideoTextMultimodality
π― What it does: This study proposes a bidirectional likelihood estimation retrieval framework (BLiM) based on a multimodal large language model (MLLM), aimed at addressing the retrieval misjudgment problem caused by candidate prior bias.
π― What it does: BokehDiff proposes a diffusion model that requires only one-step inference, utilizing a physics-inspired self-attention module to achieve direct rendering from fully focused images to realistic bokeh effects.
π― What it does: A monocular camera calibration method based on diffusion models, DM-Calib, is proposed, which can estimate camera intrinsic parameters using only a single input image and apply the results to various 3D vision tasks.
π― What it does: A residual perturbation attack (ResPA) is proposed, which utilizes the residual gradient to guide the generation of adversarial samples, enhancing transferability on black-box models.
Boosting Class Representation via Semantically Related Instances for Robust Long-Tailed Learning with Noisy Labels
Yuhang Li (Southeast University), Yuheng Jia (Southeast University)
CodeClassificationRepresentation LearningMixture of ExpertsContrastive LearningImage
π― What it does: A long-tail noisy label learning method based on instance similarity soft labels and shot-specific expert ensemble (IBC) is proposed, which dynamically adjusts label probabilities during training and targets different shot-specific experts.
π― What it does: A generative adversarial attack method named dSVA is proposed, which utilizes the dual features of self-supervised Vision Transformer (ViT) (global structural features from contrastive learning DINO and local texture features from masked image modeling MAE) to generate adversarial samples with high cross-model transferability.
Boosting MLLM Reasoning with Text-Debiased Hint-GRPO
Qihan Huang (Zhejiang University), Jie Song (Zhejiang University)
CodeLarge Language ModelReinforcement LearningPrompt EngineeringMultimodality
π― What it does: This paper improves the low data utilization and text bias issues of GRPO by introducing Hint-GRPO and text bias correction methods in MLLM inference.
π― What it does: This paper proposes SGCDet, a multi-view indoor 3D object detection framework based on adaptive 3D volume construction, which can be trained using only 3D bounding boxes, avoiding reliance on complete geometric information.
π― What it does: A separation gradient learning framework (DGL) is proposed, which eliminates the optimization conflict between the encoder and the fusion module by truncating the cross-modal loss on the encoder's gradient and replacing it with a single-modal gradient.
Bootstrap3D: Improving Multi-view Diffusion Model with Synthetic Data
Zeyi Sun (Shanghai Jiaotong University), Jiaqi Wang (Shanghai Jiaotong University)
CodeGenerationData SynthesisLarge Language ModelDiffusion modelImageVideo
π― What it does: Proposes the Bootstrap3D framework, which utilizes video diffusion models and 3D perception LLM to automatically generate high-quality multi-view images and dense descriptions, training a multi-view diffusion model.
Bootstrapping Grounded Chain-of-Thought in Multimodal LLMs for Data-Efficient Model Adaptation
Jiaer Xia (Hong Kong Baptist University), Kaiyang Zhou (Hong Kong Baptist University)
CodeDomain AdaptationData-Centric LearningTransformerLarge Language ModelVision Language ModelImageMultimodalityTabularChain-of-Thought
π― What it does: This study explores how to efficiently adapt multimodal large language models (MLLM) to specialized visual tasks with limited data (such as charts, tables, receipts, etc.) and proposes a self-verification based Grounded Chain-of-Thought (GCoT) method.
π― What it does: A cross-domain retrieval enhancement framework CRAVE is proposed, which uses image-text data to enhance malicious video detection and addresses the issue of scarce video data.
π― What it does: This study investigates the multi-organ segmentation problem and proposes a segmentation network called DGRNet based on dynamic graph reconstruction.
Bridging the Gap between Brain and Machine in Interpreting Visual Semantics: Towards Self-adaptive Brain-to-Text Decoding
Jiaxuan Chen (Zhejiang University), Gang Pan (Zhejiang University)
CodeRecognitionGenerationTransformerLarge Language ModelVision Language ModelImageTextMagnetic Resonance Imaging
π― What it does: An adaptive semantic decoding method called Mind-SA is proposed, which dynamically detects the image regions of interest in the brain and uses them as supervision to enhance the quality of brain-to-text reconstruction.