🎯 What it does: This paper proposes a 3D object detection framework DLRFusion based on 4D RADAR and LiDAR, capable of achieving robust detection under various weather conditions.
🎯 What it does: The DrUM method is proposed, utilizing core sampling, Transformer adapter (PeCA), and hierarchical softmax guidance to perform conditional-level modeling in the latent space, achieving personalized text-to-image generation without the need for fine-tuning.
🎯 What it does: Proposes a Dual Reciprocal Learning framework that collaboratively trains text-to-motion (T2M) and motion-to-text (M2T) tasks in a closed-loop manner.
Dual-Expert Consistency Model for Efficient and High-Quality Video Generation
Zhengyao Lv (Nanjing University), Ziwei Liu (Nanyang Technological University)
CodeGenerationData SynthesisComputational EfficiencyKnowledge DistillationMixture of ExpertsDiffusion modelGenerative Adversarial NetworkVideo
🎯 What it does: A parameter-efficient dual-expert consistency model (DCM) is proposed, which first trains a semantic expert for layout and motion modeling, then freezes it and adds LoRA layers for fine-tuning the detail expert, achieving high-quality video generation and efficient sampling.
Dual-Rate Dynamic Teacher for Source-Free Domain Adaptive Object Detection
Qi He (Southwest Jiaotong University), Shuai Li (Meituan Inc.)
CodeObject DetectionDomain AdaptationImage
🎯 What it does: This paper studies source-free domain adaptive object detection and proposes a Dual-Speed Dynamic Teacher (DDT) framework, which utilizes asynchronous EMA to achieve grouped updates of teacher parameters, enhancing the quality of pseudo-labels and model adaptability.
Dual-Temporal Exemplar Representation Network for Video Semantic Segmentation
Xiaolong Xu (Sichuan University), Hao Song (Sichuan University)
CodeSegmentationConvolutional Neural NetworkVideo
🎯 What it does: This paper proposes a Dual Temporal Exemplar Representation Network (DTERN) to achieve video semantic segmentation by learning local and global temporal exemplars.
🎯 What it does: A dual-dimensional adaptive joint training framework called DualReal is proposed for video customization, which can generate realistic motion while maintaining high identity consistency, addressing the identity-motion conflict problem of traditional unidimensional customization methods.
🎯 What it does: A dynamic dictionary learning framework is proposed, explicitly constructing and iteratively optimizing class ID embeddings to enhance fine-grained recognition in remote sensing image semantic segmentation.
Dynamic Multi-Layer Null Space Projection for Vision-Language Continual Learning
Borui Kang (Nanjing University), Wenbin Li (Nanjing University)
CodeTransformerVision Language ModelImageMultimodality
🎯 What it does: A dynamic multi-layer null space projection (DMNSP) strategy is proposed and implemented, specifically addressing the catastrophic forgetting problem in visual-language models during continual learning, utilizing a parameter-efficient adapter architecture and applying the projection only to the visual branch.
Dynamic-DINO: Fine-Grained Mixture of Experts Tuning for Real-time Open-Vocabulary Object Detection
Yehao Lu (Zhejiang University), Xi Li (Zhejiang University)
CodeObject DetectionTransformerMixture of ExpertsImage
🎯 What it does: Dynamic-DINO is proposed, a framework that integrates dynamic reasoning of Mixture-of-Experts (MoE) into real-time open-vocabulary object detection, capable of expanding the subnet search space while maintaining the parameter count of a single FFN;
🎯 What it does: EA-ViT is proposed, an elastic Vision Transformer framework that can generate multi-size sub-models during the adaptation phase and dynamically select them based on device resource budgets through a lightweight router.
🎯 What it does: A novel visual Mamba model for image restoration, EAMamba, is proposed, which combines a multi-head selective scanning module (MHSSM) and an all-around scanning strategy to achieve efficient long-range dependency modeling and local pixel memory retention.
🎯 What it does: A zero-shot seed selection method called ELECT is proposed, which can quickly filter out the best random seeds that maintain background consistency in instruction-driven image editing.
🎯 What it does: An efficient image editing framework named EEdit is proposed, focusing on reducing computational redundancy in image editing based on diffusion models.
Effective Training Data Synthesis for Improving MLLM Chart Understanding
Yuwei Yang (Australian National University), Liang Zheng (Ohio State University)
CodeGenerationData SynthesisLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: This paper constructs a high-quality chart training set ECD through a five-step modular process (single chart generation, composite subchart generation, visual diversification, quality filtering, QA generation), which includes over 10,000 chart images, over 300,000 question-answer pairs, covering 29 types of charts, 25 themes, and over 250 subchart combinations.
🎯 What it does: This paper proposes an AOFT strategy that generates an approximate orthogonal lower/upper projection matrix using only one learnable vector, improving the parameter-efficient fine-tuning of ViT.
Efficient Concertormer for Image Deblurring and Beyond
Pin-Hung Kuo (National Taiwan University), Ming-Hsuan Yang (University of California)
CodeRestorationTransformerImage
🎯 What it does: This paper proposes an efficient Transformer structure—Concertormer—for single-frame image deblurring, raindrop removal, JPEG compression artifact restoration, and other recovery tasks.
🎯 What it does: A lightweight pre-training scheme for event camera data, STP, is proposed, which aligns event data with image models through prompt fusion using image pre-training models.
Efficient Fine-Tuning of Large Models via Nested Low-Rank Adaptation
Lujun Li (Hong Kong University of Science and Technology), Yike Guo
CodeGenerationOptimizationComputational EfficiencyTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper presents NoRA, a nested low-rank adaptation structure that significantly reduces trainable parameters and achieves efficient fine-tuning by freezing the outer layer of LoRA and employing serial training of LoRA within it.
🎯 What it does: A lightweight multi-person human motion prediction model EMPMP is proposed, which utilizes a dual-branch (local-global) structure to extract local and global spatiotemporal features, and achieves efficient spatial-temporal feature learning through cross-layer interaction.
🎯 What it does: We propose EfficientMT, a universal motion transfer framework that converts pre-trained text-to-video diffusion models without the need for optimization during testing.
EmotiCrafter: Text-to-Emotional-Image Generation based on Valence-Arousal Model
Shengqi Dang (Tongji University), Nan Cao (Tongji University)
CodeGenerationTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: This paper proposes a Continuous Emotion Image Generation task (C‑EICG) and designs the EmotiCrafter model, which integrates Valence-Arousal (V‑A) values with free-text prompts to generate emotionally rich images.
🎯 What it does: This paper proposes a lightweight image super-resolution network called ESC, which replaces most self-attention with a convolutional attention module (ConvAttn) and utilizes Flash Attention to achieve large-window self-attention, significantly reducing computational and memory costs.
🎯 What it does: An end-to-end Associative Reasoning Network (ARN) is proposed, which simultaneously performs entity detection and relationship (triplet) prediction in videos, utilizing CLIP's semantic priors, a Predicate Association Parsing module, and a Hierarchical Attention mechanism to achieve associative reasoning of visual relationships.
Engage for All: Making Ordinary Image Descriptions Appealing Again!
Yuyan Chen, Qingpei Guo (Ant Group)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes the task of generating appealing descriptions for ordinary images, constructs a large-scale AppealImage dataset, and introduces the CharmNet framework to generate attractive descriptions.
🎯 What it does: A quaternion-based spatial-spectral interaction network, QuatPanNet, is proposed for high-resolution fusion (pansharpening) of multispectral images.
CodeAdversarial AttackConvolutional Neural NetworkTransformerLarge Language ModelImageMultimodality
🎯 What it does: Proposes a Gradient-Guided Sampling (GGS) internal iterative sampling strategy, combining MI-FGSM and other iterative attack methods to enhance the balance of black-box transferable attacks.
🎯 What it does: This paper proposes a video retrieval framework called HLFormer based on dual spaces (Euclidean + hyperbolic) to address the problem of partial relevant video retrieval (PRVR).
Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention
Shiwei Zhang (Xi'an Jiaotong University), Wei Ke (Pengcheng Laboratory)
CodeObject DetectionTransformerVision Language ModelImage
🎯 What it does: This study investigates a general strategy to enhance the accuracy of zero-shot object counting through text-guided local ranking and digit-activated global attention.
🎯 What it does: This paper proposes UniPAN, a unified distribution preprocessing strategy that uses inverse transform sampling to map pixels from different satellites to the same target distribution, thereby enhancing the generalization ability of deep learning models across different sensors.
🎯 What it does: Proposes two methods, UnionCut and UnionSeg, to reliably detect image foreground sets in unsupervised object discovery, thereby distinguishing between foreground and background and deciding when to stop discovery.
🎯 What it does: Proposes the EA6D framework, which uses a diffusion model to generate environment-independent object representations, and then predicts 6D position and orientation through a pose decoder.
🎯 What it does: This paper proposes EquiCaps, a self-supervised learning framework for pose that leverages the inherent pose awareness of capsule networks without the need for additional predictors.
ESCNet:Edge-Semantic Collaborative Network for Camouflaged Object Detection
Sheng Ye (Xiamen University), Liujuan Cao (Xiamen University)
CodeObject DetectionSegmentationTransformerImage
🎯 What it does: The ESCNet framework is proposed for camouflage object detection, utilizing dynamically coupled edge-texture perception for fine segmentation.
🎯 What it does: This paper proposes a clothing-body fitting framework ETCH based on SE(3)-equivariant tightness vectors, which maps the clothing surface to the internal body, achieving precise fitting of the body in any pose and clothing.
CodeClassificationKnowledge DistillationAdversarial AttackTransformerLarge Language ModelVision Language ModelImageBiomedical Data
🎯 What it does: A unified framework for escaping data origin verification (Escaping DOV) is proposed, where task knowledge is transferred from a teacher model to a student model using unrelated OOD datasets after learning copyright data, thereby circumventing all DOV methods.
EVDM: Event-based Real-world Video Deblurring with Mamba
Zhijing Sun (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
CodeRestorationOptical FlowImageVideo
🎯 What it does: This paper proposes an event-driven video deblurring framework EVDM based on Mamba, which can utilize long-term motion information of events to achieve high-quality video deblurring.
🎯 What it does: A dynamic scene reconstruction framework is proposed that combines event cameras with deformable 3D Gaussian splatting, significantly enhancing reconstruction and rendering quality by utilizing the ultra-high temporal and frequency information of events.
EVEv2: Improved Baselines for Encoder-Free Vision-Language Models
Haiwen Diao (Dalian University of Technology), Xinlong Wang (Beijing Academy of Artificial Intelligence)
CodeGenerationData SynthesisTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: EVEv2.0 is proposed and implemented, a novel decoder-only vision-language model without a visual encoder, along with a complete training process for visual perception from scratch.
🎯 What it does: This paper proposes to treat classification probabilities as random variables, modeling the outputs of the teacher model using a second-order Dirichlet distribution, and achieving knowledge distillation for the student model through alignment of both the expectation (macro) and the distribution itself (micro) of this distribution.
🎯 What it does: This paper proposes the I2EvDet framework, which transforms mainstream image detectors (such as RT-DETR) into object detection models capable of handling event camera data, ultimately resulting in EvRT-DETR, which achieves a record improvement in detection accuracy.
Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition
Meiqi Cao (Nanjing University of Science and Technology), Jinhui Tang (Nanjing Forestry University)
CodeRecognitionSafty and PrivacyTransformerVision Language ModelContrastive LearningVideoMultimodality
🎯 What it does: Privacy-friendly multimodal action recognition on event camera data, utilizing an event reconstruction network to generate texture-rich pseudo-RGB frames and jointly encoding them with event stacked frames.
CodeObject DetectionAnomaly DetectionGraph Neural NetworkTransformerVision Language ModelPoint Cloud
🎯 What it does: A training-free graph score propagation framework GSP is proposed for OOV detection of 3D point clouds using a visual-language model (VLM);
Exploring The Visual Feature Space for Multimodal Neural Decoding
Weihao Xia (University of Cambridge), Cengiz Oztireli (University of Cambridge)
CodeClassificationRecognitionGenerationTransformerLarge Language ModelVision Language ModelDiffusion modelMultimodalityBiomedical DataMagnetic Resonance ImagingBenchmark
🎯 What it does: Aiming at human brain fMRI signals, the VINDEX method is proposed, utilizing various visual feature spaces and a pre-trained multimodal large language model (MLLM) for zero-shot multi-level decoding of brain signals, achieving fine-grained textual descriptions, localization, and reasoning of visual stimuli.
FA: Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection
Xinhua Lu (Sun Yat-sen University), Ruixuan Wang (Sun Yat-sen University)
CodeAnomaly DetectionPrompt EngineeringVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes a Forced Prompt Learning framework based on CLIP to enhance out-of-distribution (OOD) detection performance by learning richer class descriptions under few-shot conditions.
Factorized Learning for Temporally Grounded Video-Language Models
Wenzheng Zeng (National University of Singapore), Hwee Tou Ng (National University of Singapore)
CodeTransformerVision Language ModelVideoTextBenchmark
🎯 What it does: The D2VLM framework is proposed, which separates the temporal localization of video-language models and text responses into a two-step learning process: first localization, then response.
Failure Cases Are Better Learned But Boundary Says Sorry: Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training
Yanyun Wang (Hong Kong University of Science and Technology), Li Liu (Hong Kong University of Science and Technology)
🎯 What it does: A new Robust Perception objective is proposed in adversarial training (AT), and based on this objective, the RPAT method is designed to alleviate the accuracy-robustness trade-off problem in AT.
FairGen: Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions
Yilei Jiang (Chinese University of Hong Kong), Xiangyu Yue (Chinese University of Hong Kong)
CodeGenerationData SynthesisDiffusion modelImage
🎯 What it does: A lightweight, pluggable framework called 'FairGen' is proposed, which learns attribute latent directions in diffusion models through self-supervised methods, achieving debiasing and distribution control for gender, race, and their intersectional attributes.
FairHuman: Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models
Yuxuan Wang (Beijing University of Posts and Telecommunications), Zhanyu Ma (Beijing University of Posts and Telecommunications)
CodeGenerationOptimizationDiffusion modelImage
🎯 What it does: To address the issue of distortion in facial and hand details during portrait image generation, a multi-objective fine-tuning framework called FairHuman is proposed. It constructs global and local (facial, hand) losses and achieves dynamic gradient weight allocation through the Minimum Potential Delay (MPD) fairness principle, enhancing the balance between local detail and overall quality.
FALCON: Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers
Renshan Zhang (Harbin Institute of Technology), Liqiang Nie (Huawei Noah's Ark Lab)
CodeCompressionRepresentation LearningTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: To address the issues of visual redundancy and fragmentation in high-resolution multimodal large language models (MLLM), the FALCON model is proposed, utilizing visual register technology to achieve compression and continuity in visual encoding.
Fast Globally Optimal and Geometrically Consistent 3D Shape Matching
Paul Roetzer (University of Bonn), Florian Bernard (University of Bonn)
CodeOptimizationMesh
🎯 What it does: A global optimal geometric consistent 3D shape matching method based on surface loop sets and super product graphs is proposed, solved using integer linear programming, and integer optimal solutions are obtained in experiments.
🎯 What it does: Proposes the Group Training method, which accelerates the training of 3D Gaussian Splatting through periodic grouping and caching of Gaussian atoms, while improving rendering quality.
🎯 What it does: Proposes the FastPoint software acceleration technology, which accelerates FPS and neighborhood search by predicting the distance curve between sampling points, significantly improving the inference speed of 3D point cloud models.
🎯 What it does: This paper proposes FastVAR, a post-training acceleration method that enhances the efficiency of visual autoregressive (VAR) models in high-resolution image generation by caching token pruning.
🎯 What it does: A feature dataset was constructed that includes three major types of large models (DINOv2, Llama3, SD3) across five tasks (image classification, semantic segmentation, depth estimation, common sense reasoning, text-to-image synthesis), and a unified testing condition, bit rate metric (BPFP), and benchmark evaluation process were proposed; two baseline encoders (VTM and Hyperprior) were also provided, and benchmark experiments were completed.
Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning
Zongyao Xue (Institute of Computing Technology, Chinese Academy of Sciences), Xilin Chen (Institute of Computing Technology, Chinese Academy of Sciences)
CodeClassificationRecognitionTransformerVision Language ModelContrastive LearningImage
🎯 What it does: A CLIP-based Semantic Feature Decomposition-Recombination (FDR) method is proposed for few-shot class incremental learning, addressing the issues of new class prototype bias and catastrophic forgetting.
Shuo Jin (Xi'an Jiaotong-Liverpool University), Jimin Xiao (Xi'an Jiaotong-Liverpool University)
CodeSegmentationTransformerVision Language ModelImage
🎯 What it does: This paper proposes a training-free open vocabulary semantic segmentation framework SFP, aimed at suppressing outlier propagation in CLIP attention and enhancing semantic feature representation.
FED-PsyAU: Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling
Jingting Li (Chinese Academy of Sciences), Su-Jing Wang (Chinese Academy of Sciences)
CodeRecognitionFederated LearningSafty and PrivacyGraph Neural NetworkTransformerOptical FlowImageVideo
🎯 What it does: A privacy-preserving micro-expression recognition framework (FED-PsyAU) that integrates psychological priors, dynamic AU relationships, and federated learning has been designed and implemented.
🎯 What it does: By combining diffusion models with federated learning, the FedDifRC framework is proposed, utilizing Text-Driven Diffusion Contrastive Learning (TDCL) and Noise-Driven Diffusion Consistency Regularization (NDCR) to address the issue of data heterogeneity.
Haiyang Guo (University of Chinese Academy of Sciences), Cheng-Lin Liu (University of Chinese Academy of Sciences)
CodeFederated LearningLarge Language ModelTextMultimodalityBenchmark
🎯 What it does: Proposed the FCIT benchmark and designed the DISCO framework, integrating federated learning with continuous instruction tuning to address data heterogeneity and catastrophic forgetting in LMM during distributed continuous learning.
🎯 What it does: This study proposes a model heterogeneous federated learning framework called FedRAL, which enhances model generalization and communication/computation efficiency through representation angle learning, adaptive diagonal sparsification, and element-weighted aggregation mechanisms.
🎯 What it does: A new federated learning framework, FedWSQ, has been developed, combining weight normalization and distribution-aware non-uniform quantization to enhance model convergence and communication efficiency.
FedXDS: Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning
Maximilian Andreas Hoefler (Fraunhofer Heinrich Hertz Institute), Wojciech Samek (Fraunhofer Heinrich Hertz Institute)
CodeFederated LearningSafty and PrivacyExplainability and InterpretabilityImage
🎯 What it does: FedXDS proposes a federated learning framework that utilizes XAI feature attribution for interpretable feature selection and sharing to alleviate data heterogeneity.
Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models
Xudong Li (Xiamen University), Rongrong Ji (Xiamen University)
CodeMeta LearningTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: A GRMP-IQA framework based on CLIP for meta-learning soft prompt initialization and gradient regularization is proposed for few-shot blind image quality assessment.
🎯 What it does: This paper studies the effectiveness of reducing the number of denoising steps and the inference cost per step in a post-training deployment environment without fine-tuning, and proposes the PostDiff framework that achieves training-free compression through mixed-resolution denoising and mixed module caching.
Fine-grained Spatiotemporal Grounding on Egocentric Videos
Shuo Liang (Chinese University of Hong Kong), Liwei Wang (Chinese University of Hong Kong)
CodeObject DetectionSegmentationLarge Language ModelSupervised Fine-TuningVideoTextBenchmark
🎯 What it does: The first pixel-level head perspective video localization benchmark EgoMask and its large-scale training set EgoMask-Train have been constructed, and a systematic evaluation of existing spatiotemporal localization models has been conducted.
🎯 What it does: Proposes the Transfer VAE Training (TVT) method, which migrates 8×VAE to 4×VAE and aligns with the Stable Diffusion pre-trained UNet to enhance detail preservation in Real-ISR;
Fix-CLIP: Dual-Branch Hierarchical Contrastive Learning via Synthetic Captions for Better Understanding of Long Text
Bingchao Wang (Shanghai Jiao Tong University), Wei Liu (Shanghai Jiao Tong University)
CodeGenerationRetrievalTransformerLarge Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Through dual-branch training, short texts are aligned with masked images, and long texts are aligned with original images. Learnable region prompts and unidirectional masks are designed to extract local information, and a hierarchical feature alignment module is added to enhance the understanding of long texts while retaining the performance of short texts based on CLIP.
Flash-VStream: Efficient Real-Time Understanding for Long Video Streams
Haoji Zhang (Tsinghua University), Xiaojie Jin (Beijing Jiaotong University)
CodeRecognitionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoMultimodality
🎯 What it does: This paper proposes Flash-VStream, an efficient real-time long video understanding model that utilizes a two-process asynchronous framework and Flash Memory for online processing and instant responses to long video streams.
🎯 What it does: This paper proposes the FlowTok framework, which enables the direct flow of text and images in a unified 1D token low-dimensional space, allowing for simultaneous generation from text to image and from image to text.
🎯 What it does: This paper proposes a video generation framework based on diffusion models called FontAnimate, aimed at few-shot font generation tasks.
Forgetting Through Transforming: Enabling Federated Unlearning via Class-Aware Representation Transformation
Qi Guo (Xi'an Jiaotong University), Bingyi Liu (Wuhan University of Technology)
CodeClassificationFederated 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.
🎯 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.
FREE-Merging: Fourier Transform for Efficient Model Merging
Shenghe Zheng (Harbin Institute of Technology), Hongzhi Wang (Harbin Institute of Technology)
CodeOptimizationComputational 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.
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)
CodeGenerationTransformerLarge 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;
FreeMorph: Tuning-Free Generalized Image Morphing with Diffusion Model
Yukang Cao (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
CodeGenerationData 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.
🎯 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)
CodeObject 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-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)
CodeRecognitionPose 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.
🎯 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 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)
CodeObject 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
CodeClassificationRecognitionTransformerLarge 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).
CodeRecognitionObject 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.
🎯 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.
🎯 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.
Fuzzy Contrastive Decoding to Alleviate Object Hallucination in Large Vision-Language Models
Jieun Kim (Yonsei University), Sung-Bae Cho (Yonsei University)
CodeObject 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.