CVPR 2026 Papers — Page 8
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
DarkAct: A RGB-Thermal Dataset and Fusion Framework for Multimodal Low-Light Action Recognition
Yuanjun Tan (Wuhan University), Zhigang Tu (Wuhan University)
RecognitionTransformerVideoMultimodalityBenchmark
🎯 What it does: Designed a large-scale RGB-thermal low-light action recognition dataset named DarkAct, and proposed the DarkAct-Net two-stage fusion framework, which utilizes spatiotemporal differencing, Motion-Aware Attention, and Light-Adaptive Fusion to achieve high-precision action recognition in low-light environments.
DarkShake-DVS: Event-based Human Action Recognition under Low-light and Shaking Camera Conditions
Jiaqi Chen (Beijing Institute of Technology), Liyuan Pan (Beijing Institute of Technology)
RecognitionConvolutional Neural NetworkTransformerMultimodalityBenchmark
🎯 What it does: Proposed the Event-IMU Stabilized HAR (EIS-HAR) framework, which includes IMU-based adaptive motion compensation (AIMC), iterative greedy sampling (IGS), and hybrid spatiotemporal Swin Transformer (HSTS) for human action recognition under low-light conditions and 6-DoF camera shaking.
DASH: A Meta-Attack Framework for Synthesizing Effective and Stealthy Adversarial Examples
Abdullah Al Nomaan Nafi (University of Maine), Prabuddha Chakraborty (University of Maine)
Adversarial AttackImage
🎯 What it does: Constructed a differentiable meta-attack framework called DASH, which learns weighted combinations across multiple baseline l_p attacks using soft attention and generates adversarial examples through multi-stage chained iterative processes.
Data Leakage Detection and De-duplication in Large Scale Geospatial Image Datasets
Yeshwanth Kumar Adimoolam (Cyprus University of Technology), Melinos Averkiou (CYENS Centre of Excellence)
Data-Centric LearningImageBenchmark
🎯 What it does: This paper reveals that the AICrowd dataset contains a large amount of duplicates and cross-leakage by performing perceptual hashing deduplication and leakage detection on three major remote sensing building extraction benchmark datasets, and subsequently proposes an efficient deduplication/leakage detection pipeline.
Data-Centric Meta-Learning for Robust Few-Shot Generalization
Jongmin Lim (Sungkyunkwan University), Kwangsu Kim (Sungkyunkwan University)
ClassificationData-Centric LearningMeta LearningConvolutional Neural NetworkPrompt EngineeringImageBenchmark
🎯 What it does: This paper proposes a data-centric meta-learning framework DCML that aligns task inputs through learnable visual prompts to enhance few-shot generalization performance.
Dataset Distillation by Influence Matching
Haoru Tan (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)
Knowledge DistillationImageMultimodality
🎯 What it does: This paper proposes a dataset distillation method based on Influence Matching, which directly aligns the impact of synthetic data on the final model parameters rather than the intermediate training process;
DBMSolver: A Training-free Diffusion Bridge Sampler for High-Quality Image-to-Image Translation
Sankarshana Venugopal, Jonghyun Choi (Seoul National University)
Image TranslationComputational EfficiencyDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Designed a training-free Diffusion Bridge model sampler DBMSolver for efficient image-to-image translation.
DC-Merge: Improving Model Merging with Directional Consistency
Han-Chen Zhang (Southeast University), Tong Wei (Southeast University)
Representation LearningTransformerImageTextMultimodality
🎯 What it does: This paper proposes a model fusion method called DC-Merge based on directional consistency, aiming to preserve the directional information of each task vector when merging multi-task models, thereby retaining the performance of the original tasks.
DCoAR: Deep Concept Injection into Unified Autoregressive Models for Personalized Text-to-Image Generation
Fangtai Wu (Zhejiang University), Yunlong Yu (Zhejiang University)
GenerationTransformerPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Propose a personalized text-to-image generation framework DCoAR that achieves deep concept injection on a unified autoregressive model.
DDiT: Dynamic Patch Scheduling for Efficient Diffusion Transformers
Dahye Kim (Amazon), Raghudeep Gadde (Amazon)
GenerationComputational EfficiencyTransformerDiffusion modelImageVideoTextBenchmark
🎯 What it does: Dynamically adjust the resolution (patch size) of the latent in the inference phase of diffusion transformers, significantly reducing computational costs while maintaining generation quality.
DDSF: Robust Few-Shot Learning via Disentangled Subspaces with Determinantal Point Process
Xulun Ye (Ningbo University), Jieyu Zhao (Shenzhen University)
Meta LearningTransformerDiffusion modelImage
🎯 What it does: Propose the Filter-Repair-Expand framework and DDSF, leveraging DPP to perform anomaly detection, diffusion repair, and subspace expansion on noisy support sets, constructing a robust subspace prototype;
DDT: Decoupled Diffusion Transformer
Shuai Wang (Nanjing University), Limin Wang (Nanjing University)
GenerationTransformerDiffusion modelImage
🎯 What it does: Designed and trained a diffusion Transformer (DDT) with separated low-frequency encoder and high-frequency decoder, achieving faster convergence and higher quality image generation.
DeAR: Fine-Grained VLM Adaptation by Decomposing Attention Head Roles
Yiming Ma (Chongqing Research Institute of Harbin Institute of Technology), Jianzhi Teng (Hong Kong Polytechnic University)
ClassificationDomain AdaptationRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Propose the DeAR framework, which achieves fine-grained adaptation of Vision-Language Models (VLMs) by inserting attribute word vectors into the CLIP vision encoder and text encoder, and using mask-based control of information flow based on attention head function division.
Debiased Sample Selection for Learning with Noisy Labels
Weiran Pan (Huazhong University of Science and Technology), Wenfeng Xie (Ping An Property & Casualty Insurance Company of China Limited)
ClassificationData-Centric LearningImage
🎯 What it does: This paper addresses the task of learning with noisy labels by proposing two model-structure independent modules: Marginal Distribution Adjustment (MDA) and Candidate Class Selection (CCS), to eliminate class-level and instance-level confirmation bias in small-loss selection.
Deciphering Genotype-Phenotype Mechanisms from High-Content Profiling via Knowledge-Guided Multi-modal Graph Learning
Hanjing Lin (Sun Yat-sen University), Yuedong Yang (Sun Yat-sen University)
Explainability and InterpretabilityRepresentation LearningDrug DiscoveryGraph Neural NetworkContrastive LearningImageMultimodalityTabular
🎯 What it does: Propose a knowledge-guided multi-modal graph learning framework KERNEL, integrating high-throughput cell image phenotypes with structured biological knowledge for gene regulatory network (GRN) inference, drug-target interaction (DTI) prediction, and discovery of disease subtype-specific subnetworks.
Decision Boundary-aware Generation for Long-tailed Learning
Jiacheng Yang (Xiamen University), Yang Lu (Xiamen University)
ClassificationData SynthesisConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: This paper proposes a decision boundary-aware generation framework (DBG) for long-tail learning, which analyzes the boundary ambiguity caused by head-tail migration. It designs a generator to produce information-rich samples near the boundary and combines a classifier-driven dual-branch cleaning process to remove harmful samples, thereby improving the decision space and enhancing tail class accuracy.
DeCo: Frequency-Decoupled Pixel Diffusion for End-to-End Image Generation
Zehong Ma (Peking University), Qi Tian (Huawei Inc)
GenerationTransformerDiffusion modelImageText
🎯 What it does: Propose a frequency-decoupled pixel diffusion framework DeCo, using a downsampled DiT for low-frequency semantic modeling, a lightweight pixel decoder to generate high-frequency details at full resolution, and introducing a frequency-aware flow matching loss based on JPEG quantization tables.
Decoding 3D Perception via BrainSSD: Synergistic Fusion of EEG Representations from Static and Dynamic Visual Streams
Yincheng Yao (Northwestern Polytechnical University), Shu Zhang (Northwestern Polytechnical University)
GenerationRetrievalRepresentation LearningTransformerDiffusion modelContrastive LearningImageVideoBiomedical Data
🎯 What it does: This paper proposes the BrainSSD framework, which decodes 3D visual representations and generates high-fidelity 3D object models by fusing static and dynamic EEG streams.
Decompose and Transfer: CoT-Prompting Enhanced Alignment for Open-Vocabulary Temporal Action Detection
Sa Zhu (Chinese Academy of Sciences), Bo Li (Chinese Academy of Sciences)
Object DetectionTransformerLarge Language ModelVideoTextChain-of-Thought
🎯 What it does: This paper proposes a phase-splitting and alignment-based open-vocabulary temporal action detection framework (PDA), which automatically decomposes action labels into semantic descriptions of four phases—start, middle, end, and global—via the chain-of-thought (CoT) of large language models, thereby achieving fine-grained action pattern learning and cross-category knowledge transfer.
Decompose, Mix, Adapt: A Unified Framework for Parameter-Efficient Neural Network Recombination and Compression
Nazia Tasnim (Boston University), Bryan A. Plummer (Boston University)
CompressionKnowledge DistillationRepresentation LearningImageText
🎯 What it does: Propose CRISP, a unified parameter reorganization framework that simultaneously achieves model compression (MC) and parameter-efficient fine-tuning (PEFT) by decomposing pre-trained weights into shared base matrices and learnable mixing matrices.
Deconstructing the Failure of Ideal Noise Correction: A Three-Pillar Diagnosis
Chen Feng (Queen's University Belfast), Ioannis Patras (Queen Mary University of London)
ClassificationImage
🎯 What it does: In noisy label learning (NLL), experimental validation of ideal noise correction methods (e.g., Forward Correction) reveals that even with a perfect noise transition matrix, models still experience performance collapse after training. This phenomenon is explained through macro, micro, and information-theoretic analyses, while a lightweight regularization framework (FEC/JEC) is proposed to enhance the performance of correction methods.
Decouple to Generalize: Context-First Self-Evolving Learning for Data-Scarce Vision-Language Reasoning
Tingyu Li, Cheng Tan (Shanghai Artificial Intelligence Laboratory)
Data SynthesisReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Propose the DoGe framework, achieving self-evolution of Vision-Language Models (VLM) through two-stage reinforcement learning (thinking-solving).
Decouple Your Discovery and Memory in Continual Generalized Category Discovery
Jiawei Yu (National University of Defense Technology), Kele Xu (National University of Defense Technology)
ClassificationRecognitionKnowledge DistillationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: Propose a dual-branch architecture (Discovery and Memory) for continuous generalized category discovery (C-GCD), decoupling the identification of new categories from the memorization of old categories.
Decoupled and Reusable Adaptation for Efficient Cross-Modal Transfer
Yajing Liu (Shenyang Institute of Automation, Chinese Academy of Sciences), Jiandong Tian (Shenyang Ligong University)
SegmentationDomain AdaptationTransformerPrompt EngineeringMixture of ExpertsAuto EncoderContrastive LearningImageMultimodality
🎯 What it does: Propose a two-stage cross-modal transfer paradigm: the first stage (PSST) uses unlabeled data to obtain a general modality LoRA through MAE+DINOv2 self-supervised learning; the second stage (TP-MoME) inserts visual prompts and Mixture-of-Experts into the frozen LoRA for lightweight task knowledge injection and multi-modal fusion.
Decoupled Generative Modeling for Human-Object Interaction Synthesis
Hwanhee Jung (Korea University), Sangpil Kim (Korea University)
GenerationData SynthesisTransformerVision-Language-Action ModelDiffusion modelGenerative Adversarial NetworkTextPoint CloudMeshSequentialBenchmark
🎯 What it does: Proposes a Decoupled Generative Modeling for Human-Object Interaction (DecHOI), which first uses a trajectory generator to plan 3D paths for humans and objects, then employs an action generator to synthesize fine-grained joint actions along this trajectory, and combines an adversarial discriminator to constrain distant joint contacts, achieving high-quality, manual waypoint-free human-robot interaction action generation based on text descriptions.
Decoupled Residual Denoising Diffusion Models for Unified and Data Efficient Image-to-Image Translation
Ziyue Lin (The University of Hong Kong), Liangqiong Qu (Shenyang Institute of Automation, Chinese Academy of Sciences)
Image TranslationDomain AdaptationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Proposed Decoupled Residual Denoising Diffusion Models (DRDD), achieving unified and data-efficient image-to-image translation. The method maintains domain harmony and semantic mapping by splitting the diffusion process into two stages: noise diffusion and residual diffusion.
Decoupling Bias, Aligning Distributions: Synergistic Fairness Optimization for Deepfake Detection
Feng Ding (Nanchang University), Shu Hu (Purdue University)
Anomaly DetectionExplainability and InterpretabilityConvolutional Neural NetworkVideoBenchmark
🎯 What it does: Propose a dual-mechanism collaborative optimization framework that integrates structural fairness decoupling and global distribution alignment to enhance the fairness and accuracy of deepfake detection models.
Decoupling Defense Strategies for Robust Image Watermarking
Jiahui Chen (Tsinghua University), Lifeng Sun (Tsinghua University)
Safty and PrivacyAdversarial AttackSupervised Fine-TuningImage
🎯 What it does: Designed a two-stage refinement training framework called AdvMark, which first uses specialized adversarial training to fine-tune only the encoder to maintain clean accuracy, and then enhances robustness against distortion and reconstruction attacks through direct image optimization and constraint loss, while maintaining high visual quality.
Decoupling Stability and Plasticity for Multi-Modal Test-Time Adaptation
Yongbo He (Zhejiang University), Tao Jin (Zhejiang University)
Domain AdaptationVideoMultimodalityBenchmarkAudio
🎯 What it does: Proposed a diagnostic-relief framework DASP for multi-modal test-time adaptation (MM-TTA), which can adapt to distribution drift in an unsupervised manner.
Decoupling Vision and Language: Codebook Anchored Visual Adaptation
Jason Wu (University of California, Los Angeles), Jonathan Wu (University of California, Los Angeles)
Domain AdaptationComputational EfficiencyRepresentation LearningTransformerVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Developed the CRAFT framework to fine-tune only the visual encoder in large-scale vision-language models (LVLM), mapping visual features to a stable token space through a shared discrete codebook, thereby achieving decoupling between vision and language;
DecoVLN: Decoupling Observation, Reasoning, and Correction for Vision-and-Language Navigation
Zihao Xin (Nanjing University of Aeronautics and Astronautics), Shengjun Huang (Nanjing University of Aeronautics and Astronautics)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningVision Language ModelVision-Language-Action ModelImageVideoTextMultimodality
🎯 What it does: In visual and language navigation, decoupling the observation, reasoning, and error correction processes to achieve continuous perception and closed-loop control, proposing adaptive memory refinement and error correction fine-tuning based on state-action pairs.
DeDelayed: Deleting Remote Inference Delay via On-Device Correction
Dan Jacobellis (University of Texas at Austin), Neeraja J. Yadwadkar (University of Texas at Austin)
SegmentationAutonomous DrivingComputational EfficiencyTransformerAuto EncoderVideo
🎯 What it does: Propose the DeDelayed framework, combining lightweight models on low-compute devices with high-accuracy cloud models to achieve real-time video semantic segmentation; employs delay-aware remote feature compensation for local inference.
Deep Feature Deformation Weights
Richard Liu (University of Chicago), Rana Hanocka (University of Chicago)
OptimizationComputational EfficiencyKnowledge DistillationPoint CloudMeshBenchmark
🎯 What it does: Computed weights using depth feature similarity to achieve real-time handle-based mesh deformation without optimization.
DeepAlign: Mitigating Modality Conflict through Modality-Specific Alignment
Shuo Li (Nanyang Technological University), Fei Wu (Zhejiang University)
Knowledge DistillationRepresentation LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper explores the modal conflict issue in multi-modal large language models (MLLM) and proposes the DeepAlign framework, which aligns vision and text modality-specific information through representation intervention and structure-induced distillation.
Deeper Thought, Weaker Aim: Understanding and Mitigating Perceptual Impairment during Reasoning in Multimodal Large Language Models
Ruiying Peng (Tsinghua Shenzhen International Graduate School), Xiao-Hui Li
TransformerVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: This paper proposes a training-free Visual Region Guided Attention (VRGA) framework for multi-modal large language models (MLLMs) to automatically identify visually attentive heads and reweight their attention during chain-of-thought (CoT) reasoning, prompting the model to focus on problem-related visual regions and thus alleviate perceptual decay and visual drift during reasoning.
DeepfakeImpact: A Two-Stage Benchmark with Real-World Impact in Deepfake Detection
Chaoyu Gong (Nanyang Technological University Singapore), Siqiang Luo (Nanyang Technological University Singapore)
Anomaly DetectionImageVideoBenchmark
🎯 What it does: Proposes DeepfakeImpact, a two-stage evaluation framework that incorporates both traditional technical benchmarks and social impact assessment;
DeepProtect: Proactive Face-Swapping Defense using Identity Blending and Attribute Distortion
Eungi Lee (Chonnam National University), Seok Bong Yoo (Chonnam National University)
Safty and PrivacyComputational EfficiencyPrompt EngineeringVision Language ModelGenerative Adversarial NetworkContrastive LearningImageText
🎯 What it does: Propose an active facial deepfake defense method called DeepProtect, which first dilutes facial identity features by performing identity blending in the W+ space of StyleGAN, and then suppresses the identity mapping of face-swap models without significantly degrading the original image quality by embedding adversarial watermarks generated through CLIP-based text prompts for localized attribute perturbations.
DeepScan: A Training-Free Framework for Visually Grounded Reasoning in Large Vision-Language Models
Yangfu Li, Yue Lu
Explainability and InterpretabilityComputational EfficiencyPrompt EngineeringVision Language ModelImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose a training-free framework called DeepScan, which assists large vision-language models (LVLMs) in explicitly locating, calibrating, and integrating visual evidence during inference to enhance fine-grained understanding of high-resolution images.
Defect Cue-Preserved Structural Feature Refinement for Few-Shot Anomaly Detection
Le Jiang, Si Wu (City University Of Hong Kong)
SegmentationAnomaly DetectionTransformerPrompt EngineeringVision Language ModelAuto EncoderContrastive LearningImage
🎯 What it does: Propose a structure feature refinement model called DCP-SFR based on defect clue preservation, aiming to address the gradual disappearance of defect clues during deep feature extraction in few-shot anomaly detection, and achieve high-precision anomaly localization and segmentation.
Defending Unauthorized Model Merging via Dual-Stage Weight Protection
Wei-Jia Chen (National Yang Ming Chiao Tung University), Chia-Mu Yu (National Yang Ming Chiao Tung University)
Safty and PrivacyTransformerSupervised Fine-TuningImageTextMultimodality
🎯 What it does: This paper proposes a two-stage preprocessing framework named MergeGuard, aiming to actively prevent knowledge leakage caused by unauthorized model merging through the redistribution of perturbed model weights.
Deformable Gaussian Occupancy: Decoupling Rigid and Nonrigid Motion with Factorized Distillation
Yang Gao (Ecole Polytechnique Federale De Lausanne), Alexandre Alahi (Ecole Polytechnique Federale De Lausanne)
SegmentationAutonomous DrivingKnowledge DistillationTransformerGaussian SplattingVideoPoint Cloud
🎯 What it does: Propose the DeGO framework, which achieves weakly supervised occupancy prediction in dynamic 3D scenes using a deformable Gaussian occupancy model, capable of separating rigid and non-rigid motions;
Deformation-based In-Context Learning for Point Cloud Understanding
Chengxing Lin (UESTC), Wen Li (UESTC)
RestorationSegmentationConvolutional Neural NetworkTransformerPrompt EngineeringPoint CloudBenchmark
🎯 What it does: Propose the DeformPIC framework, shifting point cloud ICL from traditional Masked Point Modeling (MPM) to a geometry deformation-based approach. It learns geometric deformation for query point clouds using example pairs (prompt pairs), enabling multi-task processing such as reconstruction, denoising, registration, and segmentation.
Degradation-Consistent Test-Time Adaptation for All-in-One Image Restoration
Ni Tang (Xiamen University), Yanyun Qu (Xiamen University)
RestorationDomain AdaptationSupervised Fine-TuningDiffusion modelImage
🎯 What it does: Propose a test-time adaptation framework called DCTTA based on degradation consistency, enabling all-in-one image restoration models to adapt online under unseen degradation distributions.
Degradation-Robust Fusion: An Efficient Degradation-Aware Diffusion Framework for Multimodal Image Fusion in Arbitrary Degradation Scenarios
Yu Shi (Hefei University of Technology), Xun Chen (University of Science and Technology of China)
RestorationGenerationDiffusion modelImageMultimodalityMagnetic Resonance ImagingPositron Emission Tomography
🎯 What it does: Propose a degradation-aware diffusion framework to achieve multi-modal image fusion under arbitrary degradation conditions.
Dehallu3D: Hallucination-Mitigated 3D Generation from a Single Image via Cyclic View Consistency Refinement
Xiwen Wang (Sichuan University), Hailun Zhang (Sichuan University)
GenerationDepth EstimationNeural Radiance FieldImagePoint CloudMesh
🎯 What it does: Proposed Dehallu3D, a framework for generating 3D meshes from a single image, leveraging dense view consistency and adaptive smoothing to eliminate hallucinations (outliers) in 3D reconstruction.
Dejavu: Towards Experience Feedback Learning for Embodied Intelligence
Shaokai Wu (Shanghai Jiao Tong University), Hongtao Lu (Shanghai Jiao Tong University)
Robotic IntelligenceReinforcement LearningVision-Language-Action ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Building upon existing Vision-Language-Action (VLA) models, we propose the Experience Feedback Network (EFN), which achieves online adaptation and learning of frozen-weight VLA models during deployment by retrieving previously executed trajectories and generating residual corrections for baseline actions.
Delta Rectified Flow Sampling for Text-to-Image Editing
Gaspard Beaudouin (Harvard University), Mengyu Wang (Harvard University)
GenerationRectified FlowImage
🎯 What it does: Proposes a Delta Rectified Flow Sampling (DRFS) framework that utilizes a differential energy function based on non-reversible, untrained velocity fields for text-guided image editing, significantly suppressing the over-smoothing phenomenon in RFDS.
DeltaQuant: 4-bit Video Diffusion Models with Spatiotemporal Delta Smoothing
Xingyang Li (Mit), Muyang Li (Nunchaku Ai)
GenerationCompressionComputational EfficiencyDiffusion modelVideo
🎯 What it does: Proposed a 4-bit weight-activation quantization method called DeltaQuant, achieving significant storage compression and acceleration for video diffusion models;
Delving Aleatoric Uncertainty in Medical Image Segmentation via Vision Foundation Models
Ruiyang Li (Xidian University), Wenping Ma (Xidian University)
SegmentationTransformerBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Leverage visual foundation models to extract features, construct a semantic-aware scale based on singular value energy distribution, thereby quantifying random uncertainty in medical image segmentation, and apply it to sample screening and dynamic loss weighting to enhance segmentation robustness.
Demo2Tutorial: From Human Experience to Multimodal Software Tutorials
Zechen Bai (National University of Singapore), Mike Zheng Shou (National University of Singapore)
SegmentationGenerationLarge Language ModelReinforcement LearningVision Language ModelVision-Language-Action ModelVideoTextMultimodality
🎯 What it does: Propose the Demo2Tutorial framework, which automatically converts raw human-computer usage demonstrations (screen recordings + operation logs) into structured, editable multimodal software tutorials.
DemoFunGrasp: Universal Dexterous Functional Grasping via Demonstration-Editing Reinforcement Learning
Chuan Mao (Peking University), Zongqing Lu (Peking University)
Robotic IntelligenceReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: Propose the DemoFunGrasp framework, achieving executable, functional grasping for arbitrary objects through demonstration editing reinforcement learning.
Den-TP: A Density-Balanced Data Curation and Evaluation Framework for Trajectory Prediction
Ruining Yang (Northeastern University), Lili Su (Northeastern University)
Autonomous DrivingData-Centric LearningTime SeriesSequential
🎯 What it does: Proposed the Den-TP framework, which debiases trajectory prediction datasets and constructs subsets through density-based partitioning, gradient-guided submodular selection, and biased sampling, while introducing density-conditioned evaluation;
DENALI: A Dataset Enabling Non-Line-of-Sight Spatial Reasoning with Low-Cost LiDARs
Nikhil Behari (Massachusetts Institute of Technology), Ramesh Raskar (Massachusetts Institute of Technology)
ClassificationObject DetectionData SynthesisConvolutional Neural NetworkTransformerPoint CloudBenchmark
🎯 What it does: Built and released the DENALI dataset, utilizing low-cost consumer LiDAR complete time-domain histograms to capture hidden object information, and achieved data-driven tasks such as non-line-of-sight (NLOS) localization, classification, and size determination based on this.
Denoise and Align: Towards Source-Free UDA for Robust Panoramic Semantic Segmentation
Yaowen Chang (Wuhan University), Zhen Dong (Wuhan University)
SegmentationDomain AdaptationTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: Propose an adaptive framework called DAPASS for panoramic semantic segmentation under source-free conditions, which significantly improves cross-domain generalization performance.
Denoising as Path Planning: Training-Free Acceleration of Diffusion Models with DPCache
Bowen Cui (Alibaba Group), Pipei Huang (Alibaba Group)
GenerationComputational EfficiencyDiffusion modelImageVideo
🎯 What it does: Proposes DPCache, a training-free diffusion model sampling acceleration framework that significantly reduces the number of sampling steps while preserving generation quality through global path planning.
Denoising, Fast and Slow: Difficulty-Aware Adaptive Sampling for Image Generation
Johannes Schusterbauer (LMU Munich), Björn Ommer (LMU Munich)
GenerationTransformerDiffusion modelFlow-based ModelImageMultimodality
🎯 What it does: This paper proposes Patch Forcing, which utilizes adaptive denoising schedules for each image patch to complete easy-to-generate regions first and provide context for difficult-to-generate regions, thereby improving the generation quality of diffusion/flow models.
Dense Metric Depth Completion from Sparse Direct Time-of-Flight Sensors
Hakyeong Kim (KAIST), Min H. Kim (KAIST)
Depth EstimationTransformerImagePoint Cloud
🎯 What it does: Propose a unified dense depth completion framework that leverages sparse direct time-of-flight (dToF) measurements with RGB images through a dual-branch Vision Transformer encoder and a lightweight decoder, achieving dense depth reconstruction under various sensors, different sparsity levels, and noise conditions.
Depth Any Endoscopy: Towards Self-Supervised Generalizable Depth Estimation in Monocular Endoscopy
Shuwei Shao (Shandong University), Zhe Min (Shandong University)
Depth EstimationConvolutional Neural NetworkMixture of ExpertsBiomedical Data
🎯 What it does: Proposed Depth Any Endoscopy (DAE), a unified self-supervised monocular endoscopy depth estimation framework capable of achieving cross-domain depth prediction in various surgical environments (e.g., laparoscopy, colonoscopy).
Depth Any Panoramas: A Foundation Model for Panoramic Depth Estimation
Xin Lin (Insta360 Research), Lu Qi (Insta360 Research)
Depth EstimationRepresentation LearningTransformerDiffusion modelGenerative Adversarial NetworkImagePoint Cloud
🎯 What it does: Proposed a panoramic metric depth foundation model named DAP, achieving large-scale, cross-domain panoramic deep learning through a three-stage pseudo-label generation process.
Depth Hypothesis Guided Iterative Refinement for Event-Image Monocular Depth Estimation
Daikun Liu (Southeast University), Changyin Sun (Southeast University)
Depth EstimationRecurrent Neural NetworkTransformerMultimodality
🎯 What it does: Proposed a HypoDepth framework for monocular depth estimation based on event-image, which converts continuous depth regression into discrete depth hypothesis search and iteratively refines the results using 3D cost volumes at multiple scales.
Depth Peeling for High-Fidelity Gaussian-Enhanced Surfel Rendering
Keyang Ye (Zhejiang University), Kun Zhou (Zhejiang University)
GenerationComputational EfficiencyGaussian SplattingImage
🎯 What it does: Proposes a depth peeling rendering method that combines two-dimensional surface points with semi-transparent boundaries and three-dimensional Gaussian distributions, enabling real-time generation of high-fidelity novel view images.
DepthFocus: Controllable Depth Estimation for See-Through Scenes
Junhong Min (Samsung Electronics), Minyong Choi (Samsung Electronics)
Depth EstimationTransformerMixture of ExpertsImage
🎯 What it does: Propose a controllable depth estimation framework called DepthFocus, which achieves hierarchical reconstruction of different depth layers in perspective scenes by inputting user-specified focal parameters.
DeRVOS: Decoupling Consistent Trajectory Generation and Multimodal Understanding for Referring Video Object Segmentation
Wenxuan Cheng (Southeast University), Wankou Yang (Southeast University)
SegmentationTransformerVision Language ModelVideoTextMultimodality
🎯 What it does: Proposed the DeRVOS framework by decomposing the RVOS task into two branches: 'trajectory consistent generation' and 'multimodal understanding', enabling parallel modeling of trajectory consistency and semantic comprehension, and achieving efficient alignment and implicit selection between the two branches through the TAIS module.
Describe Anything Anywhere At Any Moment
Nicolas Gorlo (Massachusetts Institute of Technology), Luca Carlone (Massachusetts Institute of Technology)
Object DetectionObject TrackingOptimizationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelContrastive LearningWorld ModelImageVideoTextMultimodalityPoint CloudGraphTime SeriesSequentialChain-of-Thought
🎯 What it does: Construct a real-time, scalable 4D scene graph as a spatiotemporal memory for embedded question answering and 4D reasoning.
Design Your Ad: Personalized Advertising Image and Text Generation with Unified Autoregressive Models
Yexing Xu (Sun Yat-Sen University), Yulan Guo (Sun Yat-Sen University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelAuto EncoderMultimodality
🎯 What it does: Designed and implemented a unified autoregressive model, Uni-AdGen, for one-stop generation of advertising images and text, achieving personalized ad generation through foreground awareness, instruction tuning, and coarse-to-fine prioritized preference extraction.
Designing Instance-Level Sampling Schedules via REINFORCE with James-Stein Shrinkage
Peiyu Yu (Google), Hongliang Fei (Google)
GenerationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelImageText
🎯 What it does: This paper enhances generation quality by reallocation of sampling steps in a frozen text-to-image diffusion model through one-time learning of instance-level sampling schedules (Dirichlet strategy).
Designing to Forget: Deep Semi-parametric Models for Unlearning
Amber Yijia Zheng (Purdue University), Raymond A. Yeh (Purdue University)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: This paper proposes a deep semi-parametric model (SPM) that achieves efficient and interpretable data forgetting by deleting training samples during inference;
DETACH : Decomposed Spatio-Temporal Alignment for Exocentric Video and Ambient Sensors with Staged Learning
Junho Yoon (Korea Advanced Institute of Science and Technology), Dongman Lee (Korea Advanced Institute of Science and Technology)
RecognitionContrastive LearningVideoMultimodalityTime Series
🎯 What it does: This paper proposes an unsupervised pre-training framework (DETACH) to align outdoor videos with environmental sensors, thereby enhancing the performance of non-intrusive multimodal action recognition.
DetAny4D: Detect Anything 4D Temporally in a Streaming RGB Video
Jiawei Hou (Fudan University), Jingbo Zhang (Tencent)
Object DetectionDepth EstimationTransformerVideo
🎯 What it does: Proposed an end-to-end 4D object detection framework called DetAny4D, which can real-time predict globally consistent 3D bounding boxes from RGB video streams.
Detect Any AI-Counterfeited Text Image
Chenfan Qu (South China University of Technology), Lianwen Jin (South China University of Technology)
Anomaly DetectionTransformerLarge Language ModelPrompt EngineeringDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: Constructed a large-scale AI forged text-image dataset named DanceText, and proposed DS-Net, a unified detection network, for classifying and localizing three types of forgeries in text images: full-image generation, region editing, and removal.
Detect Anything via Next Point Prediction
Qing Jiang (South China University of Technology), Lei Zhang (Peking University)
Object DetectionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodality
🎯 What it does: This paper proposes a 3B parameter multimodal large language model named Rex-Omni, which can perform object detection and various visual perception tasks by predicting the next token following coordinate words.
Detecting AI-Generated Forgeries via Iterative Manifold Deviation Amplification
Jiangling Zhang (Wuhan University of Technology), Ziyu Chen (Wuhan University of Technology)
Anomaly DetectionTransformerPrompt EngineeringAuto EncoderImage
🎯 What it does: Developed a two-stage closed-loop forgery amplification network (IFA-Net) based on a frozen MAE authenticity prior, achieving pixel-level forgery detection and localization.
Detecting Compressed AI-Generated Images via Phase Spectrum Robustness
Kai Li (Shenzhen Campus of Sun Yat-sen University), Xiaochun Cao (Shenzhen Campus of Sun Yat-sen University)
Anomaly DetectionTransformerImage
🎯 What it does: Addressing the issue of sharply declining detection performance of AI-generated images after compression on social networks
Detecting Unknown Objects via Energy-based Separation for Open World Object Detection
Jun-Woo Heo (Korea University), Gyeong-Moon Park (Korea University)
Object DetectionImageBenchmark
🎯 What it does: Proposed a new open-world object detection framework called DEUS, aiming to address the problems of incremental learning for known objects and unsupervised identification for unknown objects.
DetectSCI: Toward Object-Guided ROI Reconstruction for High-Resolution Video Snapshot Compressive Imaging
Xingjian Jiang (Zhejiang University), Xin Yuan (Westlake University)
RestorationObject DetectionCompressionTransformerVideoBenchmark
🎯 What it does: This study proposes the DetectSCI framework, which can directly detect objects in video snapshot compressive imaging (SCI) measurements and achieve object-oriented region of interest (ROI) reconstruction based on detection results, significantly reducing reconstruction computational load and memory usage.
DEVA: Fine-tuning Multimodal Large Language Models for Visual Perception Tasks
Debasmit Das (Qualcomm AI Research), Fatih Porikli (Qualcomm AI Research)
ClassificationObject DetectionSegmentationTransformerLarge Language ModelReinforcement LearningImageTextMultimodality
🎯 What it does: Fine-tune multimodal large language models using reinforcement learning to enhance performance in visual perception tasks.
DeX-Portrait: Disentangled and Expressive Portrait Animation via Explicit and Latent Motion Representations
Yuxiang Shi (University Of Science And Technology Of China), Ligang Liu (University Of Science And Technology Of China)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelGenerative Adversarial NetworkImageVideo
🎯 What it does: Propose DeX-Portrait, a facial animation method based on diffusion models, which can generate high-quality avatar animations with separable pose and expression control from a single source image and a driving video.
DextER: Language-driven Dexterous Grasp Generation with Embodied Reasoning
Junha Lee (Pohang University of Science and Technology), Minsu Cho (Pohang University of Science and Technology)
Robotic IntelligenceTransformerLarge Language ModelVision Language ModelVision-Language-Action ModelTextMultimodalityPoint CloudChain-of-Thought
🎯 What it does: Designed a contact-based embodied reasoning framework called DextER, which first predicts contact points of each link of the multi-finger fingers on the target object's surface, and then generates a complete grasping pose to achieve language-driven multi-finger grasping.
Dexterous World Models
Byungjun Kim (Seoul National University), Hanbyul Joo (Seoul National University)
GenerationData SynthesisTransformerDiffusion modelWorld ModelVideoMesh
🎯 What it does: Propose Dexterous World Models (DWM), a video diffusion-based framework that generates visual dynamics from hand interactions by using static 3D scene rendering and egocentric hand motion prediction.
DF^2-VB: Dual-level Fuzzy Fusion with View-specific Boosting for Multi-view Multi-label Classification
Yuena Lin (Beijing University of Technology), Gengyu Lyu (Beijing University of Technology)
ClassificationImageMultimodalityBiomedical Data
🎯 What it does: Propose a dual-layer fusion framework DF-VB to address the limitations of feature-level fusion and decision-level fusion in multi-view multi-label classification.
DFD-HR: Generalizable Deepfake Detection via Hierarchical Routing Learning
Jiamu Sun (Tencent Youtu Lab), Shouhong Ding (Tencent Youtu Lab)
Anomaly DetectionTransformerMixture of ExpertsVision Language ModelVideo
🎯 What it does: Proposes a hierarchical routing mechanism (Early Layer Pruning, Token Selection, and Mixture-of-Experts) implemented on visual foundation models to enhance the generalization capability of deepfake detection.
DGGT: Feedforward 4D Reconstruction of Dynamic Driving Scenes using Unposed Images
Xiaoxue Chen (AIR Tsinghua University), Hao Zhao (AIR Tsinghua University)
Autonomous DrivingTransformerDiffusion modelGaussian SplattingImageVideo
🎯 What it does: Proposes a single forward framework DGGT based on 3D Gaussian partitioning, which can reconstruct 4D dynamic driving scenes directly from sparse uncalibrated images without requiring camera pose input, providing editable outputs such as camera poses, dynamic Gaussians, and depth.
DGS: Dual Gradient and Semantic-Shift Guided Low-Rank Adaptation for Class Incremental Learning
Kai Li (East China Normal University), Ying Wen (East China Normal University)
ClassificationComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningImageBenchmark
🎯 What it does: In class-incremental learning (CIL), a novel dual-gradient and semantic drift-guided low-rank adaptation framework called DGS is proposed. It employs the LoRA module for parameter-efficient fine-tuning of pre-trained models, and balances stability and plasticity by fusing dual gradients and aligning classifiers/patch-level features, thereby suppressing catastrophic forgetting.
Diagnose, Correct, and Learn from Manipulation Failures via Visual Symbols
Xianchao Zeng (Beihang University), Yong-Lu Li (Shanghai Innovation Institute)
Anomaly DetectionData-Centric LearningRobotic IntelligenceLarge Language ModelSupervised Fine-TuningVision-Language-Action ModelVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Develop the ViFailback framework, which efficiently annotates real robot failure videos using visual symbols to generate a large-scale VQA dataset and benchmark;
Diagnosing and Repairing Unsafe Channels in Vision-Language Models via Causal Discovery and Dual-Modal Safety Subspace Projection
Jinhu Fu (Beijing University of Posts and Telecommunications), Sen Su (Beijing University of Posts and Telecommunications)
Safty and PrivacyTransformerVision Language ModelMultimodalityBenchmark
🎯 What it does: Identify key neurons and layers in VLMs causing unsafe behaviors through causal mediation analysis, and construct a dual-modal safety subspace projection method to perform safety projections on visual and textual activations during inference, thereby real-time repairing unsafe outputs.
Diagram2Structure: Unlocking LLMs' Diagram Comprehension through DiagramDiff, a Framework for Structuring Offline Diagrams
Haoxiang Hu (Institute of Software, Chinese Academy of Sciences), Hongan Wang (Cardiff University)
Image TranslationRestorationGraph Neural NetworkLarge Language ModelDiffusion modelImageGraph
🎯 What it does: Propose the DiagramDiff framework, converting offline images into structured data and enhancing LLM's capability in chart understanding and editing.
DialogueVPR: Towards Conversational Visual Place Recognition
Yukun Song (Beijing University of Posts and Telecommunications), Pengyang Wang (University of Macau)
RetrievalTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Proposed the Dialogic Visual Place Recognition (DlgPR) framework, transforming geolocation into an interactive reasoning process, and created a new dataset called DlgQuest-Cities.
DICArt: Advancing Category-level Articulated Object Pose Estimation in Discrete State-Spaces
Li Zhang (University Of Science And Technology Of China), Cewu Lu (Shanghai Jiao Tong University)
Pose EstimationDiffusion modelFlow-based ModelImage
🎯 What it does: Propose a category-level deformable object pose estimation framework called DICArt based on a discrete diffusion model.
Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs
Jinqi Luo, René Vidal (University of Pennsylvania)
Safty and PrivacyExplainability and InterpretabilityLarge Language ModelVision Language ModelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: Proposes the DACO framework, which constructs a concept dictionary using over 15,000 visual-textual concepts collected from WordNet and CC-3M, and enhances safety while maintaining general capabilities through fine-grained interventions on activations during inference of frozen multimodal large language models (MLLMs) via sparse autoencoders (SAE).
Diff-SemiER: Transparency-Aware Adaptive Fusion Diffusion Model with Generative Prior for Semi-Transparent Eyeglasses Removal
Jiahao Li (Henan University), Jingtao Guo (Henan University)
RestorationGenerationData SynthesisConvolutional Neural NetworkDiffusion modelScore-based ModelImage
🎯 What it does: Propose Diff-SemiER, a diffusion model framework for removing semi-transparent glasses, integrating generative priors and adaptive fusion.
Diff4Splat: Repurposing Video Diffusion Models for Dynamic Scene Generation
Panwang Pan, Yadong MU
GenerationData SynthesisTransformerDiffusion modelGaussian SplattingVideoPoint Cloud
🎯 What it does: This paper proposes DIFF4SPLAT, an end-to-end feed-forward framework capable of directly generating controllable dynamic 4D scenes from a single image, camera trajectory (optional text prompts), represented as deformable 3D Gaussian fields;
DiffBMP: Differentiable Rendering with Bitmap Primitives
Seongmin Hong (Seoul National University), Se Young Chun (Seoul National University)
OptimizationComputational EfficiencyGaussian SplattingImageVideo
🎯 What it does: Proposed an scalable differentiable rendering engine called DiffBMP, capable of gradient optimization for thousands of bitmap elements with parameters such as position, rotation, scaling, color, and opacity, and supports video and spatially constrained rendering.
DiffDecompose: Layer-Wise Decomposition of Alpha-Composited Images via Diffusion Transformers
Zitong Wang (Jilin University), Yiren Song (National University of Singapore)
SegmentationTransformerDiffusion modelAuto EncoderImage
🎯 What it does: Propose the LDAC task and implement the DiffDecompose framework to achieve hierarchical decomposition of a single semi-transparent/transparent image.
Differences That Matter: Auditing Models for Capability Gap Discovery and Rectification
Qihao Liu (Johns Hopkins University), Wen-Sheng Chu (Johns Hopkins University)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelReinforcement LearningDiffusion modelMultimodalityBenchmark
🎯 What it does: Proposed the AuditDM framework, which trains a multi-modal LLM auditor via reinforcement learning to automatically detect and correct model capability gaps.
Differentiable Adaptive 4D Structured Illumination for Joint Capture of Shape and Reflectance
Huakeng Ding (Zhejiang University), Hongzhi Wu (Zhejiang University)
Depth EstimationOptimizationNeural Radiance Field
🎯 What it does: Propose a differentiable adaptive 4D structured light framework that efficiently acquires the shape and reflectance properties of objects in a single shot using an LED array and an LCD mask.
Differentiable Laplacian Matrix Guided Superpixel Segmentation
Jeremy Juybari (University of Maine), Yifeng Zhu (University of Maine)
SegmentationImage
🎯 What it does: Propose a fully differentiable graph Laplacian matrix loss to directly promote superpixel connectivity during training of deep superpixel models, and define new fragmentation evaluation metrics.
Differentiable Stroke Planning with Dual Parameterization for Efficient and High-Fidelity Painting Creation
Jinfan Liu (Shanghai Jiao Tong University), Bingbing Ni (Shanghai Jiao Tong University)
GenerationOptimizationComputational EfficiencyGaussian SplattingImage
🎯 What it does: Propose a dual representation (discrete polyline and continuous Bézier curve) and implement a painting rendering framework that combines structural sensing search with gradient optimization through a differentiable bidirectional mapping, generating high-fidelity oil paintings with fewer brushstrokes.
Differentiable Vector Quantization for Rate-Distortion Optimization of Generative Image Compression
Shiyin Jiang (University of Electronic Science and Technology of China), Shuhang Gu (University of Electronic Science and Technology of China)
CompressionTransformerAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: Propose RDVQ, a differentiable vector quantization (VQ) compression framework that uses a soft distribution instead of hard quantization indices to achieve end-to-end rate-distortion (RD) optimization, and employs a self-attentive Transformer for entropy modeling during decoding.
Differentially Private 2D Human Pose Estimation
Kaushik Bhargav Sivangi (University of Glasgow), Fani Deligianni (University of Glasgow)
Pose EstimationSafty and PrivacyTransformerImageBenchmark
🎯 What it does: This paper proposes a unified differential privacy (DP) framework for 2D human pose estimation (HPE), combining subspace projection with feature-level privacy to significantly reduce the impact of noise on model performance, and systematically evaluates the method under multiple privacy budgets.
DiffGraph: An Automated Agent-driven Model Merging Framework for In-the-Wild Text-to-Image Generation
Zhuoling Li (Lancaster University), Jun Liu (Lancaster University)
GenerationGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningAgentic AIMixture of ExpertsDiffusion modelAuto EncoderImageTextGraph
🎯 What it does: Built an automated, graph-based model fusion framework called DiffGraph for dynamically integrating a large number of online expert models in text-to-image generation;