🎯 What it does: This paper proposes the DIMOS framework, which uses a dual-decomposition approach to simultaneously extract appearance and motion features from both image and event modalities, and enhances the performance of small target motion instance segmentation through multi-granularity cross-modal alignment.
🎯 What it does: Training diffusion models in pixel space using large image patches to construct global structures, supplemented by a lightweight Patch Detailer Head for local detail reconstruction;
🎯 What it does: Propose a training-agnostic open-vocabulary semantic segmentation method that does not require logit iterative optimization, directly solving the analytical solution of distribution differences to achieve pixel-level segmentation.
🎯 What it does: This paper significantly accelerates the sampling process of video diffusion transformers by training a learnable feature cache predictor combined with Restricted MeanFlow distillation, achieving a maximum speedup of 11.8× with almost no loss in generation quality.
🎯 What it does: For videos captured under time-varying colored lighting (disco lights), Gaussian Splatting is used to simultaneously perform 3D scene reconstruction and recovery of the scene's canonical (non-colored lighting) appearance, while allowing control of overall brightness during inference.
Discriminative Perception via Anchored Description for Reasoning Segmentation
Tao Yang (Northwestern Polytechnical University), Qi Wang (Northwestern Polytechnical University)
CodeSegmentationLarge Language ModelReinforcement LearningVision Language ModelContrastive LearningImageTextMultimodalityChain-of-Thought
🎯 What it does: Proposes the DPAD framework, introducing anchored descriptions and discriminative rewards in semantic segmentation tasks under reinforcement learning, enabling the model to actively distinguish objects from the background and generate more focused and concise reasoning chains;
Disentangle-then-Align: Non-Iterative Hybrid Multimodal Image Registration via Cross-Scale Feature Disentanglement
Chunlei Zhang (University of Technology Sydney), Jian Zhang (University of Technology Sydney)
CodeMultimodality
🎯 What it does: Proposes a non-iterative hybrid multimodal image registration network HRNet, which can predict rigid and non-rigid transformations simultaneously in a shared feature space.
Disentangled Textual Priors for Diffusion-based Image Super-Resolution
Lei Jiang (Nanjing University), Gangshan Wu (Nanjing University)
CodeSuper ResolutionTransformerVision Language ModelDiffusion modelMultimodality
🎯 What it does: This paper proposes DTPSR, a diffusion model-based image super-resolution framework that achieves interpretable and controllable step-by-step recovery through decoupled text priors;
DiT360: High-Fidelity Panoramic Image Generation via Hybrid Training
Haoran Feng (Insta360 Research), Lu Qi (Insta360 Research)
CodeGenerationTransformerDiffusion modelImageText
🎯 What it does: Designed and implemented the DiT360 framework, leveraging hybrid training with perspective images and panoramic images to generate high-fidelity, photorealistic 360° panoramic images.
Diversity over Uniformity: Rethinking Representation in Generated Image Detection
Qinghui He (Chongqing University of Posts and Telecommunications), Bin Xiao (Chongqing University of Posts and Telecommunications)
CodeAnomaly DetectionVision Language ModelImage
🎯 What it does: By constructing an 'anti-feature-collapse learning' framework, the method discriminates between generated images and real images, focusing on preserving diverse discriminative information to enhance robustness across different generative models.
Divide, Conquer, and Aggregate: Asymmetric Experts for Class-Imbalanced Semi-Supervised Medical Image Segmentation
Yajun Liu (Shanghai Jiao Tong University)
CodeSegmentationConvolutional Neural NetworkMixture of ExpertsBiomedical Data
🎯 What it does: This paper proposes a 'Divide, Conquer, and Aggregate' (DCA) framework to address the class imbalance problem in medical image segmentation;
🎯 What it does: Propose a single-step real-world image super-resolution framework DNF-SR, which enhances image quality by utilizing dual inputs (noisy LR + original LR) and post-training negative feature fine-tuning.
🎯 What it does: This paper proposes a feature decoupling and calibration framework (F2DC) for domain-skewed federated learning, which enhances cross-domain generalization performance by separating local features into domain-robust and domain-related components and calibrating the latter.
DPAR: Dynamic Patchification for Efficient Autoregressive Visual Generation
Divyansh Srivastava (University of California San Diego), Joshua Kimball (Dolby Laboratories)
CodeGenerationTransformerLarge Language ModelAuto EncoderImage
🎯 What it does: Propose a self-attention visual generation model named DPAR that dynamically aggregates image tokens into variable-sized patches to reduce the number of tokens and computational cost.
DPGF-Net: Dual-Prior Guided Fusion Network for Joint Assessment of Perceptual Quality and Semantic Consistency in AI-Generated Images
Tao Li (Chongqing University), Mingliang Zhou (Chongqing University)
CodeRestorationConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Proposes DPGF-Net, a dual-prior guided fusion network capable of simultaneously evaluating the perceptual quality of AI-generated images and their textual semantic consistency.
DPL: Decoupled Prototype Learning for Enhancing Robustness of Vision-Language Transformers to Missing Modalities
Jueqing Lu (Monash University), Lan Du (Monash University)
CodeClassificationRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningMultimodality
🎯 What it does: Designed and implemented a prediction head called Decoupled Prototype Learning (DPL) that adaptively handles missing modalities by separating and modality-specific decomposition of category prototypes, enhancing the robustness of Vision-Language Transformers under missing modality conditions.
DreamSR: Towards Ultra-High-Resolution Image Super-Resolution via a Receptive-Field Enhanced Diffusion Transformer
Qingji Dong (ByteDance Inc), Yitong Wang (ByteDance Inc)
CodeSuper ResolutionTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: Proposed DreamSR, a two-stage ultra-high-resolution image super-resolution model based on Diffusion Transformer, which suppresses over-generation and enhances detail reconstruction in patch-wise inference through dual-branch MM-ControlNet and Restoration Acceleration LoRA;
🎯 What it does: Propose DriveLaW, a unified framework for video generation and trajectory planning, where the latent representations from the video generator drive trajectory generation;
DrivePI: Spatial-aware 4D MLLM for Unified Autonomous Driving Understanding, Perception, Prediction and Planning
Zhe Liu (University of Hong Kong), Hengshuang Zhao (Yinwang Intelligent Technology Co. Ltd.)
CodeAutonomous DrivingTransformerLarge Language ModelVision-Language-Action ModelImageMultimodalityPoint Cloud
🎯 What it does: Propose DrivePI, a unified 4D multimodal large language model capable of simultaneously performing spatial understanding, 3D perception, prediction, and planning.
CodeGenerationReinforcement Learning from Human FeedbackTransformerReinforcement LearningDiffusion modelFlow-based ModelImageStochastic Differential Equation
🎯 What it does: Proposed a reward model based on pre-trained diffusion models (DRM), and applied it to alignment of diffusion models via reinforcement learning (Step-GRPO) and sampling improvement (Step-Sampling).
Moyang Li (Eth Zurich), Daniel Barath (Eth Zurich)
CodeAutonomous DrivingOptimizationComputational EfficiencyTransformerContrastive LearningSimultaneous Localization and MappingOptical FlowImageVideoPoint CloudBenchmark
🎯 What it does: Propose a real-time monocular dynamic SLAM system DROID-W, which estimates pixel-level dynamic uncertainty through differentiable uncertainty-aware bundle adjustment, achieving robust tracking and high-quality geometric reconstruction in dynamic environments.
Dual-Estimator: Decoupling Global and Local Semantic Shift for Drift Compensation in Class-Incremental Learning
Fankang Xu (Nanjing University of Science and Technology), Zechao Li (Nanjing University of Science and Technology)
CodeClassificationKnowledge DistillationRepresentation LearningMixture of ExpertsImageBenchmark
🎯 What it does: Propose Dual-Estimator (Dual-E) by estimating local and global semantic drift to achieve drift compensation in sample-free class incremental learning.
Dual-level Adaptation for Multi-Object Tracking: Building Test-Time Calibration from Experience and Intuition
Wen Guo (Shandong Technology and Business University), Junyu Gao (Chinese Academy of Sciences)
CodeObject TrackingTransformerVideoBenchmark
🎯 What it does: Propose a test-time calibration framework (TCEI) based on experience and intuition, enhancing identity prediction performance in multi-object tracking through two-level caches of short-term transient memory and long-term historical experience.
Kai Zhu (Wuhan University of Science and Technology), Jun Cheng (Institute for Infocomm Research)
CodeSegmentationTransformerImageBiomedical Data
🎯 What it does: Proposed a prompt-free curled structure segmentation framework (SACM) based on SAM, achieving fine-grained local adaptation and global cross-domain feature fusion through a two-layer adapter.
Dual-Level Confidence based Implicit Self-Refinement for Medical Visual Question Answering
Meihong Pan (Westlake University), Yefeng Zheng (Westlake University)
CodeTransformerLarge Language ModelContrastive LearningMultimodalityBiomedical DataBenchmark
🎯 What it does: Proposes a dual-layer confidence self-optimization framework called DuCoR, which evaluates and weights pseudo-labels by leveraging the loss distribution of pseudo-labels and feature prototype similarity, thereby achieving adaptive distribution alignment in medical visual question answering.
Dual-Level Hypergraph Generation for Addressing Feature Scarcity in Whole-Slide Image Classification
Shuilian Yao (Dalian University Of Technology), Xin Fan (Dalian University Of Technology)
CodeClassificationGraph Neural NetworkDiffusion modelAuto EncoderImageBiomedical Data
🎯 What it does: This paper proposes a dual-layer hypergraph generation framework called Dual-HGNet to address the issue of insufficient minority class features in whole-slide image classification.
Dual-Prototype-Guided Multi-task Learning for Unsupervised Anomaly Detection and Classification
Qianhao Luo (Dongguan University of Technology), Weiling Li (CISDI Group Co., Ltd)
CodeClassificationAnomaly DetectionTransformerContrastive LearningImageBiomedical Data
🎯 What it does: Propose an end-to-end multi-task learning framework PG-SFD, jointly training unsupervised anomaly detection and weakly supervised anomaly classification to address local visual ambiguity and cross-task feature conflicts.
Duala: Dual-Level Alignment of Subjects and Stimuli for Cross-Subject fMRI Decoding
Shumeng Li (Nanjing University), Yinghuan Shi (Nanjing University)
CodeRetrievalDomain AdaptationTransformerSupervised Fine-TuningVision Language ModelDiffusion modelBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposes the Duala two-layer alignment framework, which fine-tunes cross-subject visual decoding for new subjects using only one hour of fMRI data.
CodeCompressionComputational EfficiencyTransformerVision Language ModelImageVideoMultimodality
🎯 What it does: Proposes a two-stage visual token compression framework named DUET-VLM, achieving significant reduction in visual tokens while preserving visual semantics.
DVAR: Dynamic Visual Autoregressive Modeling for Image Super-Resolution
Yu Zheng (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)
CodeSuper ResolutionTransformerAuto EncoderImage
🎯 What it does: This paper proposes DVAR, a dynamic visual autoregressive image super-resolution model that can uniformly generate images at different target sizes.
Sicheng Zuo (Tsinghua University), Jiwen Lu (Tsinghua University)
CodeAutonomous DrivingTransformerSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Proposed and implemented a vision geometry Transformer (DVGT) that operates under different camera configurations without camera priors, capable of directly generating global dense 3D point maps and vehicle poses from multi-view unposed image sequences.
🎯 What it does: Propose a dynamic exposure burst image recovery (DEBIR) pipeline that achieves high-quality image reconstruction in low-light conditions by predicting the optimal exposure time for each frame, combined with BAENet and burst image recovery networks.
🎯 What it does: This study proposes the Dynamic Importance Sample Mining (DIEM) framework, integrating gradient alignment importance estimation and constrained reweighting into every optimization step of Reinforcement Learning Fine-tuning (RFT) to achieve adaptive and self-organized data usage;
🎯 What it does: Proposed a novel optimizer SGDF based on optimal linear filtering, dynamically recalibrates momentum, and real-time minimizes the mean squared error of gradient estimates.
Dynamic Token Reweighting for Robust Vision-Language Models
Tanqiu Jiang (Stony Brook University), Ting Wang (Stony Brook University)
CodeSafty and PrivacyComputational EfficiencyAdversarial AttackTransformerVision Language ModelMultimodality
🎯 What it does: Propose DTR, an inference-time defense method that resists multimodal jailbreak attacks by dynamically reweighting visual tokens in the KV cache
Xingguang Zhong (University of Bonn), Cyrill Stachniss (University of Bonn)
CodePose EstimationDepth EstimationSimultaneous Localization and MappingImageVideo
🎯 What it does: Propose a monocular visual SLAM system that integrates the depth prior of a feedforward reconstruction model and moving object segmentation, achieving robust camera pose estimation, scale-consistent dense depth reconstruction, and real-time moving object segmentation in dynamic scenes.
Dynamics-Aware Preference Optimization for Vision-Language Models
Jusheng Zhang (Sun Yat-sen University), Keze Wang (Sun Yat-sen University)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: Proposed and implemented a two-stage vision-language model preference alignment method called CW-DPO, which first smooths the loss surface through constrained SFT, and then adaptively adjusts the gradient of DPO using cooling weights based on the average token log-prob, thereby alleviating the 'compression effect' caused by easy negative samples and achieving a more stable optimization process.
Easy2Hard: From Partially to Fully Unmatched Modalities as Negative Samples in Contrastive Learning
Zhicheng Yang (Southern Illinois University), Xiaopeng Jiang (Southern Illinois University)
CodeRetrievalTransformerContrastive LearningMultimodalityBiomedical DataElectronic Health RecordsElectrocardiogram
🎯 What it does: Propose the Easy2Hard framework for multi-modal contrastive learning with M>2, which first fine-grainedly divides negative samples into two categories: partial mismatch (easy) and complete mismatch (hard). It gradually shifts the training focus from easy to hard negative samples through sigmoid curriculum scheduling, achieving more refined contrastive learning.
EasyOmnimatte: Taming Pretrained Inpainting Diffusion Models for End-to-End Video Layered Decompositio
Yihan Hu (GVC Lab, Great Bay University), Xiaodong Cun (GVC Lab, Great Bay University)
CodeSegmentationGenerationTransformerSupervised Fine-TuningMixture of ExpertsDiffusion modelVideo
🎯 What it does: Proposed a unified end-to-end video Omnimatte method that directly generates foreground layers, alpha mattes, and background layers using a fine-tuned pre-trained video inpainting diffusion model.
Echoes of Ownership: Adversarial-Guided Dual Injection for Copyright Protection in MLLMs
Chengwei Xia (Lanzhou University), Yi Yang (Zhejiang University)
CodeAdversarial AttackSupervised Fine-TuningVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Designed and implemented a trigger image method based on adversarial dual injection for tracking copyright ownership of multimodal large language models in black-box environments.
🎯 What it does: Develop a phase-aware diffusion model (EchoVDiff) capable of generating a complete cardiac cycle video from any single-frame cardiac ultrasound image.
🎯 What it does: This work proposes Edit2Perceive, a unified diffusion Transformer framework that transforms image-to-image (I2I) diffusion models into dense perception models for monocular depth estimation, surface normal estimation, and interactive matting.
Editprint: General Digital Image Forensics via Editing Fingerprint with Self-Augmentation Training
Haiwei Wu (University of Electronic Science and Technology of China), Jiantao Zhou (University of Macau)
CodeAnomaly DetectionRepresentation LearningConvolutional Neural NetworkVision Language ModelContrastive LearningImage
🎯 What it does: Designed and trained a self-supervised general image forensics feature called Editprint, which leverages an online self-incremental editing pool to simulate massive camera internal and external processing chains, thereby learning features capable of distinguishing different forensics tasks (e.g., SID, SNP, CSI).
🎯 What it does: Proposed the ConvNeur two-branch network, where one branch is responsible for local detail extraction and the other for global context aggregation, achieving efficient global-local decoupling through gated fusion;
🎯 What it does: Propose E3Flow, a SE(3)-equivariant visual motion policy integrating spherical harmonics and flow matching, achieving efficient reasoning and data efficiency for robotic manipulation tasks;
🎯 What it does: Proposed an efficient deconvolution network for solving large-scale 3D inverse problems, achieving state-of-the-art performance in 3D X-ray cone-beam CT and 3D multi-coil accelerated MRI.
🎯 What it does: Proposes an efficient framework called LAGS (Lightweight Approximation with uncertainty-adaptive Guidance Scheduling), which is training-free and can directly utilize existing pre-trained Score-based Generative Models (SGM) for weighted sampling.
EgoMind: Activating Spatial Cognition through Linguistic Reasoning in MLLMs
Zhenghao Chen (Beihang University), Di Huang (Beihang University)
CodeTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposes the EgoMind framework, which utilizes Chain-of-Thought (CoT) to achieve multi-frame spatial cognition through pure 2D language reasoning.
🎯 What it does: Propose the EDA framework, which extends the EDM general design space to arbitrary noise patterns, supports various noise covariances, and achieves image restoration without additional computational overhead while maintaining modularity;
ELV-Halluc: Benchmarking Semantic Aggregation Hallucinations in Video Understanding
Hao Lu (SenseTime Research), Lewei Lu (SenseTime Research)
CodeTransformerLarge Language ModelReinforcement LearningVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Proposed the ELV-Halluc benchmark to systematically evaluate semantic aggregation hallucination (SAH) in event-level videos, and designed adversarial question-answer pairs and the SAH Ratio metric based on this benchmark.
EMO-R3: Reflective Reinforcement Learning for Emotional Reasoning in Multimodal Large Language Models
Yiyang Fang (Wuhan University), Mang Ye (Wuhan University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Proposed and implemented the EMO-R3 framework, leveraging structured emotional thinking and reflective rewards to enhance the emotional reasoning capabilities of multimodal large language models.
🎯 What it does: Proposed the Edge-aware Multimodal Residual Diffusion Model (EMR-Diff) to fuse low-resolution HSI with high-resolution MSI, generating high-resolution HSI.
🎯 What it does: Propose a supervised learning framework based on simulating generated traces, which can extract a generic signature (GenSign) from real images for AI-generated image detection.
🎯 What it does: Propose a test-time energy adaptive framework APT without sampling or source data, treating energy as complex waves and redistributing energy using wave equations.
🎯 What it does: Achieved joint optimization of 3D Gaussian Splatting and camera pose under the condition of using only RGB images; by redesigning the camera gradient flow and introducing an image-based energy progressive alignment strategy, stable coarse-to-fine pose correction and scene reconstruction were realized.
🎯 What it does: Integrate a pre-trained single-arm manipulation strategy into dual-arm tasks using an energy-based model (EBM) to achieve dual-arm collaborative manipulation.
🎯 What it does: This paper proposes a two-stage enhanced-balanced multimodal collaboration framework (EBMC), which first performs semantic decoupling of modalities and weak modality compensation, and then achieves robustness and balance in multimodal sentiment analysis through energy-guided modality coordination and instance-aware trust distillation.
Enhancing Continual Learning of Vision-Language Models via Dynamic Prefix Weighting
Hyeonseo Jang (Yonsei University), Kibok Lee (Yonsei University)
CodeComputational EfficiencyRepresentation LearningMeta LearningTransformerVision Language ModelMultimodalityBenchmark
🎯 What it does: For the continual learning scenario of visual-language models (VLM), a dynamic prefix weighting (DPW) framework is proposed, which can adaptively assign weights to prefixes and adapters on each input token, thereby achieving efficient task adaptation while preserving pre-trained knowledge.
Enhancing Descriptive Captions with Visual Attributes for Multimodal Perception
Yanpeng Sun, Jingdong Wang
CodeClassificationRecognitionObject DetectionDepth EstimationLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Designed Cap-Workflow, an end-to-end pipeline that automatically generates image descriptions rich in fine-grained attributes and relationships by leveraging multiple visual experts and large language models.
🎯 What it does: Propose a super-resolution framework for unregistered hyperspectral images based on spectral decomposition, utilizing spectral unmixing of low-resolution HSI and unregistered high-resolution RGB reference images to achieve the fusion of spatial-spectral features.
Enhancing Video Vision Language Model with Hippocampal Sensing
Xu Cao (PediaMed AI)
CodeTransformerReinforcement LearningVision Language ModelContrastive LearningVideoTextChain-of-ThoughtAudio
🎯 What it does: Designed and trained a video vision-language model named HippoVLM, leveraging cross-modal prediction and reinforcement learning to achieve active information selection and joint reasoning.
CodeFederated LearningSafty and PrivacyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningMultimodality
🎯 What it does: This paper proposes the FedTSP method, which generates fine-grained class descriptions using large language models and constructs semantically rich global prototypes through pre-trained text models. It combines learnable prompts to achieve cross-modal alignment, thereby improving prototype quality and model performance in heterogeneous federated learning.
CodeAnomaly DetectionVision Language ModelDiffusion modelContrastive LearningImageText
🎯 What it does: Propose the Envision-Attend-Respond (EnAR) framework, which utilizes diffusion models to generate visual impressions and guides LVLM to focus on counterfactual elements in images, thereby suppressing counterfactual hallucinations in vision-language models without additional training.
🎯 What it does: This paper proposes an autoregressive diffusion model based on a single image, capable of rapidly generating diverse and physically consistent future scenes through step-by-step reasoning of sparse point trajectories.
CodeRestorationLarge Language ModelAgentic AIVision Language ModelDiffusion modelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Designed and implemented EpiAgent, an agent-centric epitaph restoration system that employs an LLM-planned Observe-Conceive-Execute-Reevaluate cycle, achieving multi-modal analysis, experience-driven tool selection, composable specialized restoration tools, and multi-perspective evaluation to recover damaged inscriptions from both visual and textual perspectives.
Erasing Thousands of Concepts: Towards Scalable and Practical Concept Erasure for Text-to-Image Diffusion Models
Hoigi Seo (Seoul National University), Se Young Chun (Seoul National University)
CodeGenerationMixture of ExpertsDiffusion modelImageText
🎯 What it does: This paper proposes an expandable concept elimination framework, ETC, which can eliminate over 2000 concepts in a single pass within text-to-image diffusion models while achieving precise removal while maintaining image quality.
EReCu: Pseudo-label Evolution Fusion and Refinement with Multi-Cue Learning for Unsupervised Camouflage Detection
Shuo Jiang (Hangzhou Dianzi University), Gang Pan (Zhejiang University)
CodeObject DetectionTransformerImageBenchmark
🎯 What it does: Proposes an unsupervised camouflaged object detection framework named EReCu based on teacher-student self-evolving pseudo-label fusion and multi-clue local perception.
🎯 What it does: Proposes ESAM++, a lightweight online 3D scene perception framework for edge devices, replacing the previously time-consuming 3D sparse UNet with a 3D sparse feature pyramid network (SFPN) to achieve multi-scale point cloud feature extraction.
EthoCLIP: Ontology-Enhanced Video-Language Pretraining for Animal Behavior Understanding
Yinuo Jing (Beijing University of Posts and Telecommunications), Zhanyu Ma (Beijing University of Posts and Telecommunications)
CodeClassificationGraph Neural NetworkTransformerLarge Language ModelVision Language ModelContrastive LearningVideoText
🎯 What it does: This paper first unifies five animal behavior datasets under the Neuro Behavior Ontology (NBO) standard, constructing the AnimalBand dataset containing 74,671 videos; subsequently, it proposes EthoCLIP, a model that integrates ontology semantics and graph attention within a visual-language contrastive learning framework.
🎯 What it does: Proposed an event-illumination collaborative low-light image enhancement framework EIC-LIE, which utilizes event and image information jointly to improve the quality of low-light images.
Ruxiao Duan (Yale University), Alex Wong (Yale University)
CodeGenerationNeural Radiance FieldImageBenchmark
🎯 What it does: Propose Evidential Neural Radiance Fields (Evidential NeRF), enabling simultaneous estimation of model uncertainty and data uncertainty with a single forward pass;
Evidential Transformation Network: Turning Pretrained Models into Evidential Models for Post-hoc Uncertainty Estimation
Yongchan Chun (Korea University), Heuiseok Lim (Korea University)
CodeExplainability and InterpretabilityComputational EfficiencyImageTextBenchmark
🎯 What it does: Propose a lightweight post-processing module called Evidential Transformation Network (ETN), which converts any pre-trained model into a reliable evidence deep learning model by learning sample-dependent affine transformations in the logit space.
🎯 What it does: Designed and implemented a lightweight Vision-Language-Action model, Evo-1, which maintains VLM semantic alignment through two-stage training, combines cross-modal diffusion transformers and integration modules, achieving end-to-end multimodal perception and continuous action generation.
🎯 What it does: Proposes the EvObj framework to achieve unsupervised 3D instance segmentation, bridging the domain gap between synthetic and real point clouds through object candidate identification and evolution modules, as well as object completion modules.
🎯 What it does: Propose an exact three-dimensional Gaussian expansion (Exact-GS) model free of approximation errors for X-ray CT reconstruction and novel view synthesis;
Shilin Xu (Sichuan University), Yuan Sun (Sichuan University)
CodeClassificationTransformerVision Language ModelMultimodality
🎯 What it does: Propose the EXOTIC framework, which leverages external visual knowledge generated by pre-trained vision-language models to guide the completion and classification of missing multi-view data, thereby improving accuracy in incomplete multi-view learning.
CodeObject DetectionExplainability and InterpretabilityImage
🎯 What it does: Propose a visualization method called VX-CODE, which leverages Shapley values and interactions to capture the collective contributions of image pixels, thereby explaining the bounding box localization and category prediction of object detectors.
🎯 What it does: Propose the Adaptive Masked Reconstruction (AMR) framework to enhance the efficiency and effectiveness of self-supervised pre-training for skeletal action recognition.
Exploring Spatiotemporal Feature Propagation for Video-Level Compressive Spectral Reconstruction: Dataset, Model and Benchmark
Lijing Cai (Nanjing University), Xun Cao (Nanjing University)
CodeRestorationTransformerVideoBenchmark
🎯 What it does: This paper constructs a high-quality dynamic hyperspectral image dataset called DynaSpec, proposes a Transformer-based video-level compressed spectral reconstruction network named PG-SVRT, and achieves significantly improved reconstruction quality and spatiotemporal consistency on SCI systems such as DD-CASSI.
Exploring the Underwater World Segmentation without Extra Training
Bingyu Li (China Telecom), Xuelong Li (China Telecom)
CodeSegmentationTransformerLarge Language ModelVision Language ModelImageTextBenchmark
🎯 What it does: Constructed a large fine-grained underwater segmentation dataset AquaOV255 and a unified evaluation benchmark UOVSBench, and proposed a training-free Earth2Ocean framework to achieve cross-domain underwater object segmentation.
🎯 What it does: Propose a single-image HDR reconstruction method called ExpoCM, which achieves high-quality HDR outputs through an exposure-aware one-step generation framework.
🎯 What it does: Propose the ExPose framework, which combines video generation models with extreme perspective pose estimation, enhancing geometric consistency by generating intermediate frames.
Exposing and Evaluating Hallucinations for GUI Grounding
Zicheng Zhang (Shanghai AI Lab), Guangtao Zhai (Shanghai AI Lab)
CodeAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmark
🎯 What it does: This paper investigates the hallucination phenomenon in graphical user interface (GUI) localization tasks, proposing two categories of hallucinations: confusion hallucinations and fabrication hallucinations, and constructing a specialized benchmark (GUI-HalluBench) to evaluate these hallucinations.
🎯 What it does: Propose a frequency-aware network called F2Net, which decomposes ultra-high-resolution remote sensing images into high- and low-frequency branches to preserve details and global semantics, respectively, achieving semantic segmentation.
🎯 What it does: Propose FSENet, a point-level weakly supervised temporal emotion localization framework, which utilizes facial features to guide multimodal interaction, and further enhances emotion boundary detection through point-aware semantic contrast and boundary smoothing pseudo labels.
🎯 What it does: FaceCam provides a portrait video generation system based on a single video and target camera trajectory, capable of precisely controlling camera position and motion while preserving subject identity and dynamic expressions.
🎯 What it does: Propose a framework based on pre-trained models, soft re-ranking retrieval, and hierarchical anchors, using pathology slides as the main benchmark. During training, missing modality information supplementation is utilized to achieve cancer risk prediction under multi-modal missing scenarios.