🎯 What it does: A two-stage knowledge distillation framework is proposed to transfer the geometric knowledge of a monocular base model to a depth completion network with sparse LiDAR input, enhancing fine-grained depth prediction.
🎯 What it does: A training-free, multi-prompt long video generation method named DiTCtrl is proposed, utilizing the MM-DiT model to achieve semantic consistency and smooth transitions between different prompts.
Divot: Diffusion Powers Video Tokenizer for Comprehension and Generation
Yuying Ge (Tencent), Ying Shan (Tencent)
CodeGenerationData SynthesisTransformerLarge Language ModelDiffusion modelVideoMultimodality
🎯 What it does: A continuous video segmenter called Divot based on diffusion models is proposed, and on this basis, Divot-LLM is constructed to unify video understanding and generation.
CodeCompressionComputational EfficiencyTransformerVision Language ModelImageVideoMultimodality
🎯 What it does: A visual token pruning method called DivPrune based on Maximum-Minimum Diversity (MMDP) is proposed, which can significantly reduce the number of tokens in large multimodal models without the need for fine-tuning or calibration sets, thereby improving inference speed and memory utilization.
🎯 What it does: This paper proposes the Cloth-Hybrid Lifelong Person Re-identification (CH-LReID) task and designs the Differentiated Knowledge Consolidation (DKC) framework to achieve dynamic knowledge balancing.
🎯 What it does: A data-free knowledge distillation method DKDM is proposed, utilizing a pre-trained diffusion model as a data source to train new diffusion models of any architecture without accessing the original training set.
🎯 What it does: A three-stage framework for removing reflections from glasses, DL2G, is proposed: first, a multiplicative degradation model is used to estimate the degraded image and obtain preliminary results; then, a local structure-aware diffusion model is employed to complete the details; finally, a global consistency refinement module integrates non-eye features from the input image to achieve reflection removal results with consistent color and lighting.
🎯 What it does: A View-Batch model is proposed to enhance the long-term memory capability of continual learning networks by adjusting the recall interval for repeated learning on the same sample.
DocLayLLM: An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding
Wenhui Liao (South China University of Technology), Lianwen Jin (South China University of Technology)
CodeTransformerLarge Language ModelTextMultimodalityChain-of-Thought
🎯 What it does: This paper presents DocLayLLM, an efficient multimodal text-rich document understanding model that achieves this by lightweight embedding of OCR results, visual patches, and two-dimensional positional information into the LLM input.
DocSAM: Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning
Xiao-Hui Li (Institute of Automation of Chinese Academy of Sciences), Cheng-Lin Liu (University of Chinese Academy of Sciences)
CodeSegmentationTransformerImageMultimodality
🎯 What it does: This paper presents DocSAM, a unified document image segmentation framework based on Transformer, capable of performing document layout analysis, multi-granularity text segmentation, and table structure recognition tasks simultaneously.
Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data
Wenxin Su (University of Shanghai for Science and Technology), Xiatian Zhu (University of Surrey)
CodeDomain AdaptationAuto EncoderImageBiomedical Data
🎯 What it does: An online model-agnostic domain adaptation (OMG-DA) framework is proposed, and cross-domain transfer for diabetic retinopathy (DR) grading is achieved based on the Generative Unseen Examples (GUES) method.
🎯 What it does: The MomAD framework is proposed, achieving temporal consistency and stability in end-to-end driving planning by introducing trajectory momentum and perceptual momentum.
🎯 What it does: The DoraCycle framework is proposed, utilizing multimodal cycles (T2I2T and I2T2I) with unaligned text and images for domain adaptation, avoiding the need for a large amount of paired data.
DPC: Dual-Prompt Collaboration for Tuning Vision-Language Models
Haoyang Li (University of Technology Sydney), Guodong Long (University of Technology Sydney)
CodeClassificationRecognitionTransformerPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Proposes the Dual-Prompt Collaboration (DPC) framework to address the Base-New Trade-off problem by decoupling the optimization of base classes and new classes at the prompt level.
🎯 What it does: A dual-pyramid recursive network DPFlow is proposed, which can adaptively handle video frame pairs from 1K to 8K without the need for high-resolution training data, and generate high-quality optical flow; at the same time, a newly established Kubric-NK dataset is released, providing dense optical flow annotations at four resolutions: 1K–8K.
CodeRecognitionObject DetectionAutonomous DrivingTransformerVision Language ModelVideoTextBenchmark
🎯 What it does: A novel traffic regulation layer task for online HD map construction is proposed, and a large-scale MapDR dataset is released, followed by two baseline methods - VLE-MEE (modular) and RuleVLM (end-to-end).
DrVideo: Document Retrieval Based Long Video Understanding
Ziyu Ma (Hunan University), Jianfei Cai (Monash University)
CodeRetrievalTransformerLarge Language ModelVision Language ModelVideoTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A long video understanding framework based on document retrieval, DrVideo, is proposed, which transforms long videos into long text documents and utilizes LLM for question answering.
DSPNet: Dual-vision Scene Perception for Robust 3D Question Answering
Jingzhou Luo (Sun Yat-sen University), Liang Lin (Sun Yat-sen University)
CodeRecognitionRetrievalTransformerVision Language ModelImageTextPoint Cloud
🎯 What it does: This paper proposes a Dual-Sight Perception Network (DSPNet) that integrates multi-view images and 3D point clouds to accomplish the 3D visual question answering (3D QA) task.
CodeSegmentationLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: The DSV-LFS framework is proposed, which combines semantic prompts generated by large language models with pixel-level matching for image support, used for few-shot semantic segmentation.
DTGBrepGen: A Novel B-rep Generative Model through Decoupling Topology and Geometry
Jing Li (University of Science and Technology of China), Falai Chen (University of Science and Technology of China)
CodeGenerationTransformerDiffusion modelMesh
🎯 What it does: A B-rep generation framework that separates topology and geometry is proposed, first generating a valid topological structure and then gradually generating geometry;
DTOS: Dynamic Time Object Sensing with Large Multimodal Model
Jirui Tian (Dalian University of Technology), Gao Huang (Dalian University of Technology)
CodeObject DetectionRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodality
🎯 What it does: A dynamic spatiotemporal target perception framework (DTOS) based on large language models is proposed to address the issues of multi-reference and visual information loss in RVOS and Moment Retrieval.
Dual Energy-Based Model with Open-World Uncertainty Estimation for Out-of-distribution Detection
Qi Chen (University of Science and Technology of China), Hu Ding (University of Science and Technology of China)
CodeClassificationAnomaly DetectionImage
🎯 What it does: A new dual energy model (DEBO) is proposed for detecting out-of-distribution samples (OOD), addressing the limitations of existing methods through a dual classifier architecture and a unified energy-based objective function.
CodeImage TranslationRestorationTransformerMixture of ExpertsVision Language ModelDiffusion modelImage
🎯 What it does: This paper proposes a dual-granularity semantic-guided diffusion model SGDiff based on sparse expert routing for achieving general full-resolution image fusion.
🎯 What it does: Designed and implemented the DualAnoDiff dual-branch diffusion model, capable of simultaneously generating anomalous images and corresponding masks, achieving high-quality, aligned anomalous image-mask pairs.
🎯 What it does: A large-scale dual-view X-ray image dataset, LDXray, is proposed, along with the Auxiliary-View Enhanced Network (AENet) dual-view detection framework, which enhances the localization accuracy of hard-to-detect categories by utilizing the cross-information between the main view and the auxiliary view, as well as expert models.
🎯 What it does: This paper proposes a weakly supervised framework for referring expression understanding, DViN, which enhances fine-grained visual description capabilities through a dynamically routable visual encoder combination.
DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models
Keda Tao (Westlake University), Huan Wang (Westlake University)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideo
🎯 What it does: This paper presents DyCoke, a training-free, pluggable dynamic visual token compression method designed to accelerate the inference of large video language models.
Dynamic Updates for Language Adaptation in Visual-Language Tracking
Xiaohai Li (Guangxi Normal University), Shuxiang Song (Guangxi Normal University)
CodeObject TrackingDomain AdaptationTransformerLarge Language ModelVision Language ModelImageVideoMultimodality
🎯 What it does: The DUTrack framework is proposed to enhance tracking robustness by dynamically updating multimodal reference information in visual-language tracking.
🎯 What it does: This paper proposes DynPose, a dynamic framework based on a lightweight router that dynamically selects lightweight or heavyweight networks for human pose estimation based on pose difficulty, significantly improving inference speed.
DynRefer: Delving into Region-level Multimodal Tasks via Dynamic Resolution
Yuzhong Zhao (University of Chinese Academy of Sciences), Fang Wan (University of Chinese Academy of Sciences)
CodeRecognitionObject DetectionTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: We propose DynRefer, a method for multi-task region-level visual language reasoning achieved through dynamic resolution (simulating human retinal focus and saccade).
🎯 What it does: This study proposes an efficient diffusion model training framework, EB-Diff-Train, which utilizes early 'early bird' sparse subnetworks. It achieves a training speed increase of over 10 times without the need to fully train a dense model, while maintaining generation quality.
CodeRobotic IntelligenceTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper presents ECBench, a comprehensive human perspective question-answering benchmark that covers static, dynamic scenes, and hallucinations, aimed at systematically evaluating the capabilities of large visual language models in embodied cognition tasks.
🎯 What it does: Proposes the EchoMimicV2 method, which generates high-quality upper-body animation videos using audio, reference images, and gesture sequences;
🎯 What it does: A unified model EchoONE has been designed and implemented, capable of performing structural segmentation of multi-plane echocardiograms within a single model, utilizing the SAM framework, PC-Mask semantic mask prompts, and LFFA local feature fusion.
Edit Away and My Face Will not Stay: Personal Biometric Defense against Malicious Generative Editing
Hanhui Wang (University of Southern California), Zhengzhong Tu (Texas A&M University)
CodeSafty and PrivacyAdversarial AttackDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes FACELOCK, a privacy protection method that eliminates facial biometric information after editing with diffusion models by generating adversarial perturbations.
Efficient Depth Estimation for Unstable Stereo Camera Systems on AR Glasses
Yongfan Liu (University of California), Hyoukjun Kwon (University of California)
CodeDepth EstimationImage
🎯 What it does: Achieve low-latency, real-time stereo depth estimation on AR glasses, proposing two models: MultiHeadDepth (optimized cost volume) and HomoDepth (eliminating preprocessing).
🎯 What it does: Designed and implemented an asynchronous UNet network RDG for real-time rendering, utilizing decoupled G-buffer guidance to achieve detail-rich and temporally stable video super-resolution;
EfficientLLaVA: Generalizable Auto-Pruning for Large Vision-language Models
Yinan Liang (Tsinghua University), Jiwen Lu (Tsinghua University)
CodeOptimizationComputational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: This paper proposes EfficientLLaVA, an automatic pruning method for large visual-language models (LVLM), which achieves efficient inference with a minimal number of proxy samples without significantly losing task performance.
🎯 What it does: A lightweight visual network called EfficientViM is proposed, which efficiently captures global context using Hidden State Mixing with State Space Duality (HSM-SSD) and enhances representation capability through multi-stage hidden state fusion.
🎯 What it does: This paper proposes an efficient 3D decoder, EffiDec3D, aimed at significantly reducing the parameter count and computational requirements of 3D medical image segmentation models.
Effortless Active Labeling for Long-Term Test-Time Adaptation
Guowei Wang (South China University of Technology), Changxing Ding (South China University of Technology)
CodeDomain AdaptationImage
🎯 What it does: An adaptive method called EATTA for long-term testing is proposed, which requires labeling at most one sample per batch, utilizes feature perturbation to identify boundary samples, and achieves a balance between supervised and unsupervised objectives through dynamic weighting of gradient norms.
EgoTextVQA: Towards Egocentric Scene-Text Aware Video Question Answering
Sheng Zhou (University of Science and Technology of China), Angela Yao (National University of Singapore)
CodeAutonomous DrivingTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: The EgoTextVQA dataset has been constructed, containing 1.5K perspective videos and 7K real scene text-related QA pairs, focusing on outdoor driving and indoor housekeeping tasks.
EmoDubber: Towards High Quality and Emotion Controllable Movie Dubbing
Gaoxiang Cong (Institute of Computing Technology, Chinese Academy of Sciences), Qingming Huang (Institute of Computing Technology, Chinese Academy of Sciences)
🎯 What it does: A controllable emotional movie dubbing model called EmoDubber is proposed, addressing three major challenges: audio-video synchronization, clear pronunciation, and controllable emotions.
Yiyang Fang (Wuhan University), Mang Ye (Wuhan University)
CodeRecognitionKnowledge DistillationTransformerMixture of ExpertsMultimodality
🎯 What it does: A dynamic modality weight allocation and unimodal distillation multimodal emotion recognition framework EMOE based on Mixture of Experts is proposed.
🎯 What it does: The EDF (Emphasize Discriminative Features) method is proposed, which utilizes Grad-CAM activation mapping to dynamically enhance discriminative regions and employs Common Pattern Dropout to eliminate low-loss gradients, thereby improving dataset distillation; at the same time, a CompDD benchmark aimed at complex scenes is constructed.
🎯 What it does: A text-based vector graphic generation framework called Dream3DVG is proposed, which allows for arbitrary viewpoints and consistent visibility dependent on the viewpoint.
🎯 What it does: A training-independent encapsulated video synthesis framework EVS is proposed, which can simultaneously improve the image quality and motion consistency of text-to-video (T2V) generation.
🎯 What it does: A multi-frame attack framework based on LiDAR, OMP-Attack, is proposed, which can continuously induce trajectory prediction errors on multi-frame historical trajectories.
Enhanced OoD Detection through Cross-Modal Alignment of Multi-Modal Representations
Jeonghyeon Kim (Seoul National University of Science and Technology), Sangheum Hwang (Seoul National University of Science and Technology)
CodeClassificationAnomaly DetectionSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a Cross-Modal Alignment (CMA) multimodal fine-tuning method for multimodal visual language models (such as CLIP) to enhance out-of-domain detection (OoDD) and ID classification performance.
🎯 What it does: A weakly supervised 3D gaze estimation framework ST-WSGE is proposed, and a modality-agnostic Gaze Transformer (GaT) model is designed to achieve unified training and inference for images and videos.
🎯 What it does: This paper studies a black-box attack method that generates transferable adversarial examples using checkpoints from a single model training process.
Enhancing Creative Generation on Stable Diffusion-based Models
Jiyeon Han (Korea Advanced Institute of Science and Technology), Jaesik Choi (Korea Advanced Institute of Science and Technology)
CodeGenerationDiffusion modelImage
🎯 What it does: A training-independent C3 method is proposed, which enhances creative generation by amplifying low-frequency features in the shallow blocks of the U-Net in Stable Diffusion.
Enhancing Facial Privacy Protection via Weakening Diffusion Purification
Ali Salar (University of Oulu), Guoying Zhao (University of Oulu)
CodeGenerationSafty and PrivacyDiffusion modelImage
🎯 What it does: This paper proposes a method for adversarial modification in latent space using diffusion models, introducing learnable non-text embeddings and self-attention constraints to generate facial privacy protection images that effectively conceal the original identity while maintaining visual quality.
Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization
Sihao Liu (Harbin Institute of Technology), Bernard Ghanem (King Abdullah University of Science and Technology)
CodeClassificationOptimizationMixture of ExpertsContrastive LearningImage
🎯 What it does: This paper proposes a pluggable module S6MOD, which can significantly enhance the adaptability of online continual learning (OCL) and prevent catastrophic forgetting by adding branches based on the Selective State Space Model (S6) while keeping the original network structure unchanged.
EntropyMark: Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection
Ming Sun (Institute of Information Engineering, Chinese Academy of Sciences), Yuanfang Guo (Beihang University)
CodeClassificationObject DetectionSafty and PrivacyConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A harmless reverse gate watermark implemented through entropy constraints—EntropyMark, is proposed for copyright protection of open-source datasets.
EquiPose: Exploiting Permutation Equivariance for Relative Camera Pose Estimation
Yuzhen Liu (Chinese Academy of Sciences), Qiulei Dong (Chinese Academy of Sciences)
CodePose EstimationPoint Cloud
🎯 What it does: A general framework called EquiPose is proposed, which can enforce pose permutation equivariance (PPE) for any end-to-end relative camera pose estimation model and enhance its performance.
🎯 What it does: This paper proposes ERUPT, a model capable of learning implicit camera poses and performing efficient view synthesis with only a small number of uncalibrated images.
Tae-Young Lee (Korea University), Gyeong-Moon Park (Korea University)
CodeSafty and PrivacyComputational EfficiencyConvolutional Neural NetworkTransformerReinforcement LearningImage
🎯 What it does: This paper proposes a Knowledge Deletion (KD) framework aimed at achieving the complete elimination of specified knowledge in trained models, meeting users' privacy and practical needs.
🎯 What it does: The first method for arbitrary point tracking (ETAP) using a pure event camera is proposed, capable of achieving robust tracking in high dynamic range and high motion speed scenarios.
Evaluating Model Perception of Color Illusions in Photorealistic Scenes
Lingjun Mao (University of California), Alane Suhr (University of California)
CodeGenerationData SynthesisTransformerVision Language ModelImageChain-of-Thought
🎯 What it does: This study investigates the perception of color illusions by visual language models (VLM) in real-world scenarios and constructs a large-scale Realistic Color Illusion Dataset (RCID).
🎯 What it does: A new test-time domain adaptation framework (SDA) is proposed, which utilizes diffusion models to generate synthetic domains and fine-tune the source model, while mapping target data to the same synthetic domain to achieve domain alignment for cross-domain TTA.
🎯 What it does: This paper proposes the EvOcc framework, which utilizes evidence theory to construct a 3D semantic occupancy map from partially labeled LiDAR data, and uses this high-quality occupancy map to supervise the occupancy prediction model of multi-view cameras.
Explaining Domain Shifts in Language: Concept Erasing for Interpretable Image Classification
Zequn Zeng (Xidian University), Jiawei Ma (City University of Hong Kong)
CodeDomain AdaptationExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImage
🎯 What it does: This paper proposes a language-guided concept elimination framework (LanCE) to enhance the generalization ability of concept bottleneck models in cross-domain tasks by removing the influence of domain-specific concepts.
🎯 What it does: A TV3S architecture based on state space models is proposed, which achieves efficient video semantic segmentation by independently and parallelly processing spatial blocks and incorporating optional gating and window shifting.
🎯 What it does: This paper proposes a method that utilizes an automated exploration agent to generate interactive data in a large number of virtual game environments, and based on this, trains a transferable multi-environment world model (GenieRedux-G).
Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation
Zhiwei Yang (Fudan University), Zhijian Song (Fudan University)
CodeSegmentationTransformerLarge Language ModelVision Language ModelContrastive LearningImage
🎯 What it does: Proposes ExCEL, which generates CAM through patch-text alignment of CLIP, enhancing weakly supervised semantic segmentation performance.
Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection
Wei Luo (Tsinghua University), Wenyong Yu (Huazhong University of Science and Technology)
CodeAnomaly DetectionTransformerImage
🎯 What it does: A novel anomaly detection framework called INP-Former based on Vision Transformer is proposed, which utilizes the 'Intrinsic Normal Prototypes (INPs)' within a single test image for adaptive reconstruction, enabling anomaly detection and localization without relying on external normal samples.
🎯 What it does: A semi-supervised LiDAR semantic segmentation method based on scene affinity (AIScene) is proposed, which combines a teacher-student framework with pseudo-labels for learning.
🎯 What it does: A Semantic Feature Discrimination (SFD) framework is proposed, utilizing pixel-level semantic features from CLIP and text-guided global features for adversarial discrimination, thereby enhancing the perceptual quality of super-resolution (SR) and directly reusing the discriminator in no-reference image quality assessment (OU NR-IQA) to achieve high-accuracy unlabelled quality evaluation.
🎯 What it does: This paper proposes a three-stage Extrapolating and Decoupling framework to improve motion controllability and motion amplitude in image-to-video (I2V) generation, achieved through a lightweight adapter, training-free extrapolation, and parameter decoupling.
CodeObject DetectionSegmentationConvolutional Neural NetworkTransformerVision Language ModelImageMultimodality
🎯 What it does: Utilizing the frozen large multimodal model (LMM) word-image attention maps, a lightweight CNN mask decoder and SAM mask refiners are constructed to achieve visual alignment without fine-tuning the LMM parameters.
Face Forgery Video Detection via Temporal Forgery Cue Unraveling
Zonghui Guo (Ocean University of China), Shiguang Shan (Chinese Academy of Sciences)
CodeAnomaly DetectionTransformerVideo
🎯 What it does: A facial forgery video detection framework TFCU is proposed, which progressively excavates temporal forgery clues (instantaneous anomalies, gradual inconsistencies, cumulative distortions) to enhance detection robustness.
FaceBench: A Multi-View Multi-Level Facial Attribute VQA Dataset for Benchmarking Face Perception MLLMs
Xiaoqin Wang (Shenzhen University), Linlin Shen (Shenzhen University)
CodeRecognitionTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This paper presents the FaceBench dataset and the Face-LLaVA baseline model for evaluating the capabilities of multimodal large language models in facial attribute recognition and question answering.
FactCheXcker: Mitigating Measurement Hallucinations in Chest X-ray Report Generation Models
Alice Heiman (Stanford University), Pranav Rajpurkar (Harvard University)
CodeObject DetectionGenerationTransformerLarge Language ModelPrompt EngineeringImageTextBiomedical DataComputed Tomography
🎯 What it does: The FactCheXcker framework is proposed to correct measurement hallucinations in chest X-ray reports through a query-code-update process.
Fancy123: One Image to High-Quality 3D Mesh Generation via Plug-and-Play Deformation
Qiao Yu (Huazhong University of Science and Technology), Min Chen (South China University of Technology)
CodeGenerationDiffusion modelImageMesh
🎯 What it does: This paper presents a framework named Fancy123, which achieves the generation of high-quality 3D meshes from a single image, focusing on enhancing multi-view consistency and fidelity of the input image through 2D and 3D deformation modules, and significantly improving texture clarity through a 'projection' operation.
Fast and Accurate Gigapixel Pathological Image Classification with Hierarchical Distillation Multi-Instance Learning
Jiuyang Dong (Harbin Institute of Technology), Yongbing Zhang (Harbin Institute of Technology)
CodeClassificationComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: A hierarchical distillation multi-instance learning framework (HDMIL) is proposed, which first trains a self-distillation dynamic multi-instance network (DMIN) using high-resolution WSI to generate patch importance masks, and then trains a lightweight pre-screening network (LIPN) using low-resolution WSI, significantly reducing inference time while maintaining or even improving classification performance.
🎯 What it does: For multi-frame LiDAR point clouds, we propose FASTer, a lightweight long-term 3D object detection framework that dynamically acquires focus tokens and scales them within the transformer.
FedBiP: Heterogeneous One-Shot Federated Learning with Personalized Latent Diffusion Models
Haokun Chen (Siemens Technology), Volker Tresp (Ludwig Maximilian University of Munich)
CodeData SynthesisFederated LearningDiffusion modelImageBiomedical Data
🎯 What it does: A heterogeneous one-round federated learning framework FedBiP based on a dual-layer personalized diffusion model is proposed to address data scarcity and feature heterogeneity issues.
FedSPA: Generalizable Federated Graph Learning under Homophily Heterogeneity
Zihan Tan (Wuhan University), Mang Ye (Wuhan University)
CodeFederated LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes the FedSPA framework, which achieves federated graph learning under homogeneity and heterogeneity through subgraph feature propagation separation and homogeneity bias-driven aggregation.
🎯 What it does: This paper proposes the Few-shot Implicit Function Generation task, which aims to train a generative model using only a few target category INRs to produce diverse and functionally consistent weight samples.
FFR: Frequency Feature Rectification for Weakly Supervised Semantic Segmentation
Ziqian Yang (XJTLU University of Liverpool), Jimin Xiao (XJTLU University of Liverpool)
CodeSegmentationTransformerImage
🎯 What it does: Proposes the Frequency Feature Rectification (FFR) framework to correct the boundary mis-segmentation caused by the attenuation of high-frequency features in Vision Transformer-based weakly supervised semantic segmentation, thereby improving pseudo-label and segmentation accuracy.
FG^2: Fine-Grained Cross-View Localization by Fine-Grained Feature Matching
Zimin Xia (Ecole Polytechnique Federale de Lausanne), Alexandre Alahi (Ecole Polytechnique Federale de Lausanne)
CodeObject DetectionPose EstimationContrastive LearningSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: This paper proposes a fine-grained cross-view localization method that maps ground image features to 3D point clouds and learns feature selection in the height dimension, thereby generating a BEV plane corresponding to aerial images and matching it to estimate the 3DoF pose of the ground camera.
🎯 What it does: A post-training quantization method based on Fisher Information Matrix (FIM) approximation, called FIMA-Q, is proposed, specifically targeting block-level reconstruction quantization for Vision Transformers.
Finding Local Diffusion Schrodinger Bridge using Kolmogorov-Arnold Network
Xingyu Qiu (Harbin Institute of Technology), Shuo Li (Harbin Institute of Technology)
CodeGenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper studies a framework for finding Local Schrödinger Bridges (LDSB) in the diffusion path subspace, significantly improving image generation quality and sampling efficiency by optimizing the diffusion path weights (fA(t), fB(t)) using a pre-trained denoising network.
Finer-CAM: Spotting the Difference Reveals Finer Details for Visual Explanation
Ziheng Zhang (Ohio State University), Wei-Lun Chao (Ohio State University)
CodeExplainability and InterpretabilityImageMultimodality
🎯 What it does: The Finer-CAM method is proposed, which generates fine-grained saliency maps by comparing the target class with similar classes, thereby improving the detail localization ability of visual explanations and being compatible with existing CAM methods such as Grad-CAM and Score-CAM; this idea is also extended to multimodal zero-shot models.
CodeRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningVideo
🎯 What it does: This paper constructs the FineVD database and proposes the FineVQ model for multi-dimensional fine-grained quality assessment of UGC videos.
🎯 What it does: This paper proposes the FiRe framework, which combines a general restoration model with its degradation operations during training, using its fixed points as implicit priors to solve linear inverse problems.
FLAIR: VLM with Fine-grained Language-informed Image Representations
Rui Xiao (Technical University of Munich), Stephan Alaniz (Technical University of Munich)
CodeSegmentationRetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningImageText
🎯 What it does: Designed and trained a visual-language model named FLAIR, which learns fine-grained image representations on local sub-regions of images through text-conditioned attention pooling, achieving more precise image-text alignment.
FLAME: Frozen Large Language Models Enable Data-Efficient Language-Image Pre-training
Anjia Cao (Xi'an Jiaotong University), Zhiheng Ma (Chinese Academy of Sciences)
CodeRetrievalComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: The paper proposes the FLAME framework, which utilizes a frozen large language model (LLM) as a text encoder to directly process long texts and achieve efficient language-image pre-training.
🎯 What it does: In the 3D point cloud Transformer, the authors propose the Flash3D Transformer, which unifies geometric locality with GPU memory locality. By utilizing Perfect Spatial Hashing (PSH) to map points to contiguous memory and implementing Bucket-and-Swin window shifting without additional cost, it significantly enhances computational efficiency.
🎯 What it does: This paper presents FlashGS, a high-performance 3D Gaussian splatting rendering framework that significantly enhances the real-time rendering speed of large-scale high-resolution scenes.
FLAVC: Learned Video Compression with Feature Level Attention
Chun Zhang (Waseda University), Jiro Katto (Waseda University)
CodeCompressionTransformerVideo
🎯 What it does: A Transformer-based video compression framework called FLAVC is proposed, which incorporates a feature-level attention module, a dense overlapping patcher, and a Transformer-CNN hybrid encoder.
🎯 What it does: The FlexUOD framework is proposed to automatically estimate contamination factors during unsupervised image anomaly detection and select appropriate anomaly detection methods based on the signal-to-noise ratio.