CVPR 2025 Papers — Page 7
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2871 papers
DiN: Diffusion Model for Robust Medical VQA with Semantic Noisy Labels
Erjian Guo (University of Sydney), Luping Zhou (University of Sydney)
TransformerDiffusion modelBiomedical Data
🎯 What it does: Proposes a noise label robust framework DiN for medical visual question answering.
Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection
Jia Guo (Tsinghua University), Hongen Liao (Tsinghua University)
Anomaly DetectionTransformerImage
🎯 What it does: This paper proposes Dinomaly, a multi-class unsupervised anomaly detection framework based on pure Transformer, addressing the performance disadvantages of a unified model in multi-class scenarios.
DINOv2 Meets Text: A Unified Framework for Image- and Pixel-Level Vision-Language Alignment
Cijo Jose (Meta), Piotr Bojanowski (Meta)
ClassificationSegmentationRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This study presents dino.txt, a framework that aligns the self-supervised visual foundation model DINOv2 with language, achieving strong performance in zero-shot classification and open vocabulary semantic segmentation.
DIO: Decomposable Implicit 4D Occupancy-Flow World Model
Christopher Diehl (TU Dortmund University), Raquel Urtasun (University of Toronto)
Autonomous DrivingWorld ModelOptical FlowPoint Cloud
🎯 What it does: A decodable implicit 4D occupancy-flow world model DIO is proposed, which can learn scene occupancy, instance occupancy, and flow prediction from sparse LiDAR observations, and supports instance prompting at any time.
Directional Label Diffusion Model for Learning from Noisy Labels
Senyu Hou (Shanxi University), Wenjian Wang (Shanxi University)
ClassificationDiffusion modelContrastive LearningImage
🎯 What it does: Designed and trained a dual-channel directional label diffusion model to learn high-quality classifications from noisy labels.
DirectTriGS: Triplane-based Gaussian Splatting Field Representation for 3D Generation
Xiaoliang Ju (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
GenerationData SynthesisDiffusion modelAuto EncoderGaussian SplattingPoint Cloud
🎯 What it does: The DirectTriGS framework is constructed, using triplane representation for Gaussian Splatting (GS) and decoding it into point clouds through a differentiable TriRenderer, followed by the direct generation of text to 3D GS using VAE and a two-stage diffusion model.
DiSciPLE: Learning Interpretable Programs for Scientific Visual Discovery
Utkarsh Mall (Columbia University), Carl Vondrick (Columbia University)
Explainability and InterpretabilityComputational EfficiencyLarge Language ModelImage
🎯 What it does: A framework called DiSciPLE is proposed, which utilizes LLM and evolutionary search to automatically generate interpretable programs, focusing on scientific visual discovery tasks.
Disco4D: Disentangled 4D Human Generation and Animation from a Single Image
Hui En Pang (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
GenerationData SynthesisPose EstimationDiffusion modelGaussian SplattingImage
🎯 What it does: Disco4D generates animatable 4D human models from a single image and separates the body from clothing using Gaussian Splatting;
Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models
Yan Xie (Xidian University), Hongwei Liu (Xidian University)
ClassificationObject DetectionExplainability and InterpretabilityTransformerContrastive LearningImage
🎯 What it does: A concept bottleneck model based on decomposed optimal transport (DOT-CBM) is designed and implemented for fine-grained image-concept association and enhanced interpretability.
Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning
Xueyi Ke (Nanyang Technological University), Bihan Wen (Nanyang Technological University)
ClassificationRecognitionContrastive LearningImageText
🎯 What it does: This paper studies whether the CVCL model trained with infant perspectives and parental speech can learn visual concepts beyond its linguistic vocabulary range, and verifies the model's ability to recognize previously unseen objects through neuron labeling and NeuronClassifier.
DiscoVLA: Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval
Leqi Shen (Tsinghua University), Guiguang Ding (Tsinghua University)
RetrievalKnowledge DistillationTransformerVision Language ModelContrastive LearningVideoText
🎯 What it does: This paper proposes DiscoVLA, a parameter-efficient CLIP adaptation framework for video-text retrieval, addressing three types of discrepancies: visual, linguistic, and alignment.
Discrete to Continuous: Generating Smooth Transition Poses from Sign Language Observations
Shengeng Tang (Hefei University of Technology), Richang Hong (Hefei University of Technology)
GenerationData SynthesisPose EstimationTransformerDiffusion modelVideo
🎯 What it does: Generate continuous sign language videos by filling in smooth transition frames between discrete sign segments to achieve semantically coherent sign language flow.
Disentangled Pose and Appearance Guidance for Multi-Pose Generation
Tengfei Xiao, Wenping Ma
GenerationPose EstimationTransformerDiffusion modelImage
🎯 What it does: A multi-pose generation framework is proposed, decoupling pose control from appearance generation to achieve high-quality generation of multiple target poses.
Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality
Ramchandran Muthukumar (Johns Hopkins University), Rene Vidal (University of Pennsylvania)
Safty and PrivacyAdversarial AttackImage
🎯 What it does: A local variable direction projection displacement threat model based on observed training data is proposed, which can distinguish between safe and unsafe image perturbations.
DiskVPS: Vanishing Point Detector via Hough Transform in a Disk Region
Jianping Wu (Suzhou Vocational University)
Object DetectionComputational EfficiencyImage
🎯 What it does: A Hough transform method called DiskVPS is proposed, which maps the image plane to a disk-shaped space and uses single edge voting for efficient and robust vanishing point detection.
Dispider: Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction
Rui Qian (Chinese University of Hong Kong), Jiaqi Wang (Shanghai Innovation Institute)
GenerationOptimizationTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: Designed and implemented Dispider, a decoupled framework for real-time interactive video large language models, supporting continuous monitoring, timely decision-making, and asynchronous responses;
DiSRT-In-Bed: Diffusion-Based Sim-to-Real Transfer Framework for In-Bed Human Mesh Recovery
Jing Gao (Carnegie Mellon University), Zackory Erickson (Carnegie Mellon University)
Pose EstimationDepth EstimationDomain AdaptationDiffusion modelImageMultimodalityMesh
🎯 What it does: This paper proposes a simulation-to-real transfer framework based on diffusion models, utilizing large-scale synthetic depth images and limited or no real data for training, thereby achieving 3D mesh recovery of human bodies on beds.
Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability
Yingdong Shi (ShanghaiTech University), Kan Ren (ShanghaiTech University)
GenerationExplainability and InterpretabilityConvolutional Neural NetworkDiffusion modelAuto EncoderImage
🎯 What it does: This paper proposes a mechanism interpretable method called DIFFLENS to identify and eliminate social biases in diffusion models, particularly biases related to attributes such as gender, age, and race.
Distilled Prompt Learning for Incomplete Multimodal Survival Prediction
Yingxue Xu (Hong Kong University of Science and Technology), Hao Chen (Hong Kong University of Science and Technology)
Knowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringMultimodalityBiomedical Data
🎯 What it does: A Distilled Prompt Learning (DisPro) framework is proposed, using a two-stage prompting method to address the issue of missing modalities in multimodal survival prediction, leveraging large language models to complete the missing information.
Distilling Long-tailed Datasets
Zhenghao Zhao (University of Illinois Chicago), Yan Yan (National University of Singapore)
Knowledge DistillationData-Centric LearningImage
🎯 What it does: This paper proposes the Dataset Distillation for Long-Tail Data (LTDD) task, which can compress imbalanced raw data into a small-scale, uniformly distributed synthetic dataset.
Distilling Monocular Foundation Model for Fine-grained Depth Completion
Yingping Liang (Beijing Institute of Technology), Ying Fu (Beijing Institute of Technology)
RestorationDepth EstimationKnowledge DistillationImagePoint Cloud
🎯 What it does: A two-stage knowledge distillation framework is proposed to transfer the geometric knowledge of a monocular base model to a depth completion network with sparse LiDAR input, enhancing fine-grained depth prediction.
Distilling Multi-modal Large Language Models for Autonomous Driving
Deepti Hegde (Johns Hopkins University), Fatih Porikli (Qualcomm)
Autonomous DrivingKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningMultimodality
🎯 What it does: Injecting world knowledge from multimodal large language models into a visual end-to-end driving planner through a distillation method to achieve efficient and robust path planning.
Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment
Xudong Li (Xiamen University), Liujuan Cao (Xiamen University)
Knowledge DistillationTransformerContrastive LearningImage
🎯 What it does: A Vision Transformer-based BIQA framework called SHDIQA is proposed, which achieves the perception and assessment of spatial heterogeneous distortions in images through block-level distortion modeling, local-global feature aggregation, and knowledge distillation.
Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation
Chanyoung Kim (Yonsei University), Seong Jae Hwang (Yonsei University)
SegmentationKnowledge DistillationTransformerContrastive LearningImage
🎯 What it does: This paper proposes a training-free open-vocabulary semantic segmentation model called CASS, which distills the spectral information from a visual foundation model into the attention of CLIP and utilizes object presence priors to improve text embeddings, enhancing the unified segmentation of different object parts.
DistinctAD: Distinctive Audio Description Generation in Contexts
Bo Fang (City University of Hong Kong), Antoni B. Chan (City University of Hong Kong)
GenerationTransformerLarge Language ModelContrastive LearningVideoTextAudio
🎯 What it does: Proposes the DistinctAD two-stage framework for generating differentiated audio descriptions (AD) that align visual and textual elements in long videos while avoiding contextual redundancy.
Distinguish Then Exploit: Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment
Weiming Liu (Bytedance Inc), Lianyong Qi (China University of Petroleum East China)
Domain AdaptationSupervised Fine-TuningImage
🎯 What it does: This paper proposes a source-side unsupervised open set domain adaptation method named DTE (Distinguish Then Exploit), which aims to efficiently transfer domains in situations where unknown category samples exist in the target domain and source data cannot be accessed. It first estimates the distinction between known and unknown target samples through weight barcode estimation, and then assigns reliable pseudo-labels to known samples through sparse label allocation.
Distraction is All You Need for Multimodal Large Language Model Jailbreaking
Zuopeng Yang (Guangzhou University), Changyu Dong (Guangzhou University)
RetrievalAdversarial AttackTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A multi-modal large language model (MLLM) jailbreak framework based on visual interference, named CS-DJ, is proposed, combining query decomposition and dual interference of comparative sub-images.
Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection
Fuyun Wang (Nanjing University of Science and Technology), Zhen Cui (Beijing Normal University)
Anomaly DetectionConvolutional Neural NetworkDiffusion modelImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A framework for open-set supervised anomaly detection based on Distribution Prototype Diffusion Learning (DPDL) is designed and implemented. It utilizes the Schrodinger bridge to map normal samples into a multi-Gaussian prototype space and employs the hyperspherical vMF distribution for feature dispersion learning, enhancing the discrimination and generalization ability for unknown anomalies.
DiTASK: Multi-Task Fine-Tuning with Diffeomorphic Transformations
Krishna Sri Ipsit Mantri (Purdue University), Moshe Eliasof (University of Cambridge)
TransformerSupervised Fine-TuningImage
🎯 What it does: The DITASK method is proposed, which achieves parameter-efficient fine-tuning in multi-task learning by maintaining the singular vectors of the ViT pre-trained weights and performing reversible continuous transformations on the singular values.
DiTCtrl: Exploring Attention Control in Multi-Modal Diffusion Transformer for Tuning-Free Multi-Prompt Longer Video Generation
Minghong Cai (Chinese University of Hong Kong), Xiangyu Yue (Chinese University of Hong Kong)
GenerationData SynthesisTransformerDiffusion modelVideoTextMultimodalityBenchmark
🎯 What it does: A training-free, multi-prompt long video generation method named DiTCtrl is proposed, utilizing the MM-DiT model to achieve semantic consistency and smooth transitions between different prompts.
DIV-FF: Dynamic Image-Video Feature Fields For Environment Understanding in Egocentric Videos
Lorenzo Mur-Labadia (Universidad de Zaragoza), Ruben Martinez-Cantin (Universidad de Zaragoza)
Object DetectionSegmentationNeural Radiance FieldContrastive LearningVideoMultimodality
🎯 What it does: Proposes the DIV-FF framework, which decomposes scene geometry and semantics in first-person videos using three streams (persistent, dynamic, actor), and integrates CLIP image-language features and EgoVideo video-language features to achieve fine segmentation and consistent reasoning of dynamic objects and affordances.
DiverseFlow: Sample-Efficient Diverse Mode Coverage in Flows
Mashrur M. Morshed (Michigan State University), Vishnu Boddeti (Michigan State University)
GenerationData SynthesisDiffusion modelFlow-based ModelRectified FlowImageText
🎯 What it does: This paper proposes DiverseFlow, a sampling method driven by DPP during flow model inference, achieving diversified outputs under a limited sampling budget.
Divide and Conquer: Heterogeneous Noise Integration for Diffusion-based Adversarial Purification
Gaozheng Pei (University of Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
RestorationAdversarial AttackDiffusion modelImage
🎯 What it does: A heterogeneous noise forward process and a dual-stage denoising diffusion model based on attention are proposed for the removal of adversarial perturbations.
Divot: Diffusion Powers Video Tokenizer for Comprehension and Generation
Yuying Ge (Tencent), Ying Shan (Tencent)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelVideoMultimodality
🎯 What it does: A continuous video segmenter called Divot based on diffusion models is proposed, and on this basis, Divot-LLM is constructed to unify video understanding and generation.
DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models
Saeed Ranjbar Alvar (Huawei Technologies Canada), Yong Zhang (Huawei Technologies Canada)
CompressionComputational EfficiencyTransformerVision Language ModelImageVideoMultimodality
🎯 What it does: A visual token pruning method called DivPrune based on Maximum-Minimum Diversity (MMDP) is proposed, which can significantly reduce the number of tokens in large multimodal models without the need for fine-tuning or calibration sets, thereby improving inference speed and memory utilization.
DKC: Differentiated Knowledge Consolidation for Cloth-Hybrid Lifelong Person Re-identification
Zhenyu Cui (Peking University), Yuxin Peng (Peking University)
RecognitionRetrievalKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes the Cloth-Hybrid Lifelong Person Re-identification (CH-LReID) task and designs the Differentiated Knowledge Consolidation (DKC) framework to achieve dynamic knowledge balancing.
DKDM: Data-Free Knowledge Distillation for Diffusion Models with Any Architecture
Qianlong Xiang (Harbin Institute of Technology), Liqiang Nie (Illinois Institute of Technology)
GenerationKnowledge DistillationConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: A data-free knowledge distillation method DKDM is proposed, utilizing a pre-trained diffusion model as a data source to train new diffusion models of any architecture without accessing the original training set.
DL2G: Degradation-guided Local-to-Global Restoration for Eyeglass Reflection Removal
Zhilv Yi (Wuhan University), Chunxia Xiao (Hunan Normal University)
RestorationConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: A three-stage framework for removing reflections from glasses, DL2G, is proposed: first, a multiplicative degradation model is used to estimate the degraded image and obtain preliminary results; then, a local structure-aware diffusion model is employed to complete the details; finally, a global consistency refinement module integrates non-eye features from the input image to achieve reflection removal results with consistent color and lighting.
DNF: Unconditional 4D Generation with Dictionary-based Neural Fields
Xinyi Zhang (Technical University of Munich), Angela Dai
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: A 4D shape representation and Transformer diffusion model based on dictionary learning is proposed for unconditional generation of high-quality 4D animations.
DnLUT: Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables
Sidi Yang (University of Hong Kong), Ngai Wong (University of Hong Kong)
RestorationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a lightweight color image denoising framework based on lookup tables (DnLUT), achieving high-quality denoising.
Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System?
Yancheng Cai (University of Cambridge), Rafal Mantiuk (University of Cambridge)
Representation LearningContrastive LearningImage
🎯 What it does: To investigate whether foundational computer vision models (such as DINO, DINOv2, OpenCLIP, etc.) can produce responses similar to the human visual system in low-level visual tasks (contrast detection, masking, contrast constancy), a set of evaluation protocols consisting of 9 psychophysical experiments was designed, and 45 models were tested.
Do ImageNet-trained Models Learn Shortcuts? The Impact of Frequency Shortcuts on Generalization
Shunxin Wang (University of Twente), Nicola Strisciuglio (University of Twente)
ClassificationDomain AdaptationAdversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes an efficient Hierarchical Frequency Shortcut Search (HFSS) method for identifying frequency shortcuts learned in large-scale image classification models (such as ImageNet) and evaluating the impact of these shortcuts on model generalization and robustness in different out-of-distribution (OOD) scenarios.
Do Visual Imaginations Improve Vision-and-Language Navigation Agents?
Akhil Perincherry (Oregon State University), Stefan Lee (Oregon State University)
GenerationData SynthesisTransformerDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes the use of visual imagination generated by text-to-image diffusion models as an additional modality in the visual language navigation task, enhancing navigation performance through alignment loss.
Do We Always Need the Simplicity Bias? Looking for Optimal Inductive Biases in the Wild
Damien Teney (Idiap Research Institute), Ehsan Abbasnejad (University of Adelaide)
ClassificationMeta LearningImageTabular
🎯 What it does: This paper adjusts the simplicity bias of neural networks through a meta-learning-based Spline activation function without prior assumptions, and validates its advantages across various tasks.
Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models?
Yanbo Wang (University of Chinese Academy of Sciences), Ran He (Chinese Academy of Sciences)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodality
🎯 What it does: This study investigates the safety alignment issue of multimodal large language models (MLLMs) and explores whether carefully curated malicious data is needed to enhance safety.
Do Your Best and Get Enough Rest for Continual Learning
Hankyul Kang (Ajou University), Jongbin Ryu (Ajou University)
Representation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: A View-Batch model is proposed to enhance the long-term memory capability of continual learning networks by adjusting the recall interval for repeated learning on the same sample.
DocLayLLM: An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding
Wenhui Liao (South China University of Technology), Lianwen Jin (South China University of Technology)
TransformerLarge Language ModelTextMultimodalityChain-of-Thought
🎯 What it does: This paper presents DocLayLLM, an efficient multimodal text-rich document understanding model that achieves this by lightweight embedding of OCR results, visual patches, and two-dimensional positional information into the LLM input.
Docopilot: Improving Multimodal Models for Document-Level Understanding
Yuchen Duan (Shanghai AI Laboratory), Wenhai Wang (Shanghai AI Laboratory)
TransformerLarge Language ModelVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: A large-scale high-quality document-level multimodal question-answering dataset, Doc-750K, has been proposed, and a local multimodal model, Docopilot, dependent on no retrieval (RAG) has been trained based on this dataset, achieving efficient understanding and reasoning of multi-page documents.
DocSAM: Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning
Xiao-Hui Li (Institute of Automation of Chinese Academy of Sciences), Cheng-Lin Liu (University of Chinese Academy of Sciences)
SegmentationTransformerImageMultimodality
🎯 What it does: This paper presents DocSAM, a unified document image segmentation framework based on Transformer, capable of performing document layout analysis, multi-granularity text segmentation, and table structure recognition tasks simultaneously.
Document Haystacks: Vision-Language Reasoning Over Piles of 1000+ Documents
Jun Chen (King Abdullah University of Science and Technology), Mohamed Elhoseiny (King Abdullah University of Science and Technology)
RetrievalTransformerLarge Language ModelVision Language ModelTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes two retrieval question-answering benchmarks, DocHaystack and InfoHaystack, aimed at a scale of thousands of documents, and introduces the V-RAG visual retrieval-augmented generation framework, which enhances retrieval and reasoning performance using a multimodal visual encoder and LMM filter.
DocVLM: Make Your VLM an Efficient Reader
Mor Shpigel Nacson (AWS AI Labs), Ron Litman
RecognitionCompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: A universal OCR module DocVLM is designed for visual language models (VLM), which integrates OCR text and layout information into VLM through query compression, significantly improving document understanding performance under low-resolution inputs.
DoF-Gaussian: Controllable Depth-of-Field for 3D Gaussian Splatting
Liao Shen (Huazhong University of Science and Technology), Chen Change Loy (Nanyang Technological University)
RestorationData SynthesisDepth EstimationGaussian SplattingImage
🎯 What it does: A controllable depth-of-field method called DoF-Gaussian based on 3D Gaussian Splatting has been developed, which can reconstruct clear scenes from shallow depth-of-field images and supports arbitrary depth control.
DOF-GS: Adjustable Depth-of-Field 3D Gaussian Splatting for Post-Capture Refocusing, Defocus Rendering and Blur Removal
Yujie Wang (Peking University), Baoquan Chen (Peking University)
RestorationGenerationGaussian SplattingImage
🎯 What it does: A model based on 3D Gaussian splatting has been implemented, incorporating a finite aperture camera model and differentiable depth of field rendering, enabling adjustable focus points and depth of field control after capture.
Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data
Wenxin Su (University of Shanghai for Science and Technology), Xiatian Zhu (University of Surrey)
Domain AdaptationAuto EncoderImageBiomedical Data
🎯 What it does: An online model-agnostic domain adaptation (OMG-DA) framework is proposed, and cross-domain transfer for diabetic retinopathy (DR) grading is achieved based on the Generative Unseen Examples (GUES) method.
Domain Generalization in CLIP via Learning with Diverse Text Prompts
Changsong Wen (Shanghai Jiao Tong University), Wei Shen (Shanghai Jiao Tong University)
Domain AdaptationRepresentation LearningTransformerPrompt EngineeringContrastive LearningImageText
🎯 What it does: By learning diverse text prompts to enhance the performance of CLIP in domain generalization tasks, three techniques are proposed: domain-sensitive channel suppression, cross-domain consistency constraints, and orthogonal uniformization.
Don't Shake the Wheel: Momentum-Aware Planning in End-to-End Autonomous Driving
Ziying Song (Beijing Jiaotong University), Yadan Luo (University of Queensland)
Autonomous DrivingRecurrent Neural NetworkReinforcement LearningMultimodality
🎯 What it does: The MomAD framework is proposed, achieving temporal consistency and stability in end-to-end driving planning by introducing trajectory momentum and perceptual momentum.
Doppelgangers and Adversarial Vulnerability
George Kamberov (University of Alaska Anchorage)
Adversarial Attack
🎯 What it does: This paper introduces the concept of 'Adversarial Doppelgangers' (AD) and systematically analyzes the susceptibility and robustness of machine learning classifiers to such attacks from the perspective of perceptual topology.
Doppelgangers++: Improved Visual Disambiguation with Geometric 3D Features
Yuanbo Xiangli (Cornell University), Noah Snavely (Cornell University)
ClassificationRecognitionTransformerImage
🎯 What it does: By improving the image-based binary classifier, this paper addresses 3D reconstruction errors caused by visual confusions (doppelgangers) and seamlessly integrates the classifier into the SfM process.
Dora: Sampling and Benchmarking for 3D Shape Variational Auto-Encoders
Rui Chen (ByteDance Seed), Ping Tan (ByteDance Seed)
GenerationData SynthesisTransformerAuto EncoderMeshBenchmark
🎯 What it does: This paper presents Dora-VAE, a 3D VAE that utilizes Sharp Edge Sampling and a Dual Cross Attention mechanism for high-quality, low-dimensional shape encoding and reconstruction, and based on this, constructs the Dora-Bench evaluation benchmark.
DoraCycle: Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles
Rui Zhao (Show Lab National University of Singapore), Mike Zheng Shou (Show Lab National University of Singapore)
GenerationDomain AdaptationTransformerSupervised Fine-TuningImageTextMultimodality
🎯 What it does: The DoraCycle framework is proposed, utilizing multimodal cycles (T2I2T and I2T2I) with unaligned text and images for domain adaptation, avoiding the need for a large amount of paired data.
DORNet: A Degradation Oriented and Regularized Network for Blind Depth Super-Resolution
Zhengxue Wang (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)
RestorationDepth EstimationSuper ResolutionConvolutional Neural NetworkMixture of ExpertsContrastive LearningImageMultimodality
🎯 What it does: This paper proposes a deblurring method for real-world scene depth super-resolution called DORNet, which utilizes self-supervised degradation representation learning and routing-based degradation regularization to achieve adaptive modeling of unknown degradations, and restores high-resolution depth maps through degradation-guided feature transformation for RGB-D fusion.
DPC: Dual-Prompt Collaboration for Tuning Vision-Language Models
Haoyang Li (University of Technology Sydney), Guodong Long (University of Technology Sydney)
ClassificationRecognitionTransformerPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Proposes the Dual-Prompt Collaboration (DPC) framework to address the Base-New Trade-off problem by decoupling the optimization of base classes and new classes at the prompt level.
DPFlow: Adaptive Optical Flow Estimation with a Dual-Pyramid Framework
Henrique Morimitsu (University of Science and Technology Beijing), Xu-Cheng Yin (University of Science and Technology Beijing)
Image TranslationRecurrent Neural NetworkOptical FlowVideo
🎯 What it does: A dual-pyramid recursive network DPFlow is proposed, which can adaptively handle video frame pairs from 1K to 8K without the need for high-resolution training data, and generate high-quality optical flow; at the same time, a newly established Kubric-NK dataset is released, providing dense optical flow annotations at four resolutions: 1K–8K.
DPSeg: Dual-Prompt Cost Volume Learning for Open-Vocabulary Semantic Segmentation
Ziyu Zhao (University of South Carolina), Song Wang (Shenzhen University of Advanced Technology)
SegmentationConvolutional Neural NetworkPrompt EngineeringDiffusion modelImage
🎯 What it does: In open vocabulary semantic segmentation, a dual-prompt cost-based learning method that combines visual prompts and text prompts is proposed.
DPU: Dynamic Prototype Updating for Multimodal Out-of-Distribution Detection
Shawn Li (University of Southern California), Yue Zhao (University of Southern California)
Anomaly DetectionContrastive LearningVideoMultimodality
🎯 What it does: A DPU (Dynamic Prototype Updating) framework is proposed to address the issue of prediction inconsistency caused by intra-class variation in multi-modal OOD detection.
Dr. Splat: Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration
Kim Jun-Seong (POSTECH), Tae-Hyun Oh (POSTECH)
Object DetectionSegmentationLarge Language ModelGaussian SplattingPoint Cloud
🎯 What it does: A method is proposed to directly register CLIP language embeddings into a 3D Gaussian Splitting model, achieving open vocabulary 3D scene understanding.
Dragin3D: Image Editing by Dragging in 3D Space
Weiran Guang (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
Image TranslationGenerationDiffusion modelImagePoint Cloud
🎯 What it does: An interactive image editing method called Dragin3D is designed to achieve object rotation in three-dimensional space by dragging dual points.
DRAWER: Digital Reconstruction and Articulation With Environment Realism
Hongchi Xia (University of Illinois Urbana-Champaign), Wei-Chiu Ma (Cornell University)
GenerationRobotic IntelligenceLarge Language ModelDiffusion modelVideo
🎯 What it does: This paper proposes the DRAWER framework, which can automatically convert a single indoor scene video into an interactive, physically dynamic high-fidelity digital twin environment.
DreamCache: Finetuning-Free Lightweight Personalized Image Generation via Feature Caching
Emanuele Aiello (Politecnico di Torino), Enrico Magli (Samsung R&D Institute UK)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: This paper proposes DreamCache, a lightweight personalized image generation method that does not require fine-tuning, utilizing feature caching and lightweight attention adapters to achieve dynamic modulation from reference images to generated images.
DreamOmni: Unified Image Generation and Editing
Bin Xia (Chinese University of Hong Kong), Jiaya Jia
GenerationData SynthesisTransformerVision Language ModelDiffusion modelRectified FlowImageMultimodality
🎯 What it does: This paper presents DreamOmni, a unified image generation and editing model capable of simultaneously performing multiple tasks such as text-to-image generation, instructive editing, inpainting/extending, drag-and-drop editing, and reference image generation.
DreamRelation: Bridging Customization and Relation Generation
Qingyu Shi (Peking University), Xiangtai Li (Nanyang Technological University)
GenerationPose EstimationTransformerPrompt EngineeringDiffusion modelImageBenchmark
🎯 What it does: The DreamRelation framework is proposed to generate accurate object relationships based on text prompts while maintaining the identity of user-provided image objects. It includes three main modules: a data engine, keypoint matching loss (KML), and local token injection to achieve relationship-aware customized generation.
DreamText: High Fidelity Scene Text Synthesis
Yibin Wang (Fudan University), Cheng Jin (Fudan University)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: A high-fidelity scene text synthesis method based on diffusion models, called DREAMTEXT, has been developed. It uses character-level guided attention correction and jointly trains the text encoder and generator to achieve precise text rendering.
DreamTrack: Dreaming the Future for Multimodal Visual Object Tracking
Mingzhe Guo (Beijing Jiaotong University), Zhipeng Zhang (Shanghai Jiaotong University)
Object TrackingTransformerVideoMultimodality
🎯 What it does: A DreamTrack framework based on history to future is designed, utilizing the Future Dreaming module to predict future environments and enhance the generalization ability of visual target tracking through multimodal trajectory output.
DRiVE: Diffusion-based Rigging Empowers Generation of Versatile and Expressive Characters
Mingze Sun (Tsinghua University), Ruqi Huang (Tsinghua University)
GenerationData SynthesisDiffusion modelGaussian SplattingPoint CloudMesh
🎯 What it does: Designed and implemented the DRiVE framework, which can generate 3D Gaussian characters from a single image or text, and automatically complete rigging for details such as skeletons, clothing, and hairstyles, achieving high-quality and controllable animation rendering.
DriveDreamer4D: World Models Are Effective Data Machines for 4D Driving Scene Representation
Guosheng Zhao (Chinese Academy of Sciences), Xingang Wang (Chinese Academy of Sciences)
Data SynthesisAutonomous DrivingDiffusion modelGaussian SplattingWorld ModelVideoPoint Cloud
🎯 What it does: By first generating multi-view synthetic videos with complex maneuvers using a driving world model, and then aligning them with real data to train a 4D Gaussian point cloud model, the quality of driving scene reconstruction in closed-loop simulation is improved.
DriveGEN: Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation
Hongbin Lin, Zhen Li
Object DetectionDomain AdaptationAutonomous DrivingDiffusion modelImage
🎯 What it does: This paper proposes a training-free controllable text-image diffusion method called DriveGEN for enhancing the robustness of vision-centered 3D object detection models in out-of-domain (OOD) scenarios.
DriveGPT4-V2: Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving
Zhenhua Xu (University of Hong Kong), Hengshuang Zhao (University of Hong Kong)
Autonomous DrivingTransformerLarge Language ModelReinforcement LearningMultimodality
🎯 What it does: This paper proposes DriveGPT4-V2, which utilizes a multimodal large language model to achieve closed-loop end-to-end autonomous driving, combined with an expert LLM for online imitation learning.
DriveScape: High-Resolution Driving Video Generation by Multi-View Feature Fusion
Wei Wu (Sensetime Research), Chenjing Ding (Sensetime Research)
GenerationData SynthesisAutonomous DrivingTransformerDiffusion modelVideoMultimodality
🎯 What it does: This paper presents DriveScape, an end-to-end high-resolution driving video generation framework based on multi-view feature fusion, supporting sparse conditional control.
Driving by the Rules: A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map
Xinyuan Chang (Amap Alibaba Group), Xing Wei (Xi'an Jiaotong University)
RecognitionObject DetectionAutonomous DrivingTransformerVision Language ModelVideoTextBenchmark
🎯 What it does: A novel traffic regulation layer task for online HD map construction is proposed, and a large-scale MapDR dataset is released, followed by two baseline methods - VLE-MEE (modular) and RuleVLM (end-to-end).
DrivingSphere: Building a High-fidelity 4D World for Closed-loop Simulation
Tianyi Yan (University of Macau), Jianbing Shen (University of Macau)
GenerationAutonomous DrivingDiffusion modelVideo
🎯 What it does: We propose DrivingSphere, a generative closed-loop simulation framework based on 4D occupancy grids and diffusion models, capable of constructing high-fidelity 4D worlds and generating multi-view, spatiotemporally consistent video outputs for the evaluation and validation of end-to-end autonomous driving algorithms.
DroneSplat: 3D Gaussian Splatting for Robust 3D Reconstruction from In-the-Wild Drone Imagery
Jiadong Tang (Beijing Institute of Technology), Yi Yang (Hangzhou Dianzi University)
RestorationSegmentationOptimizationGaussian SplattingImagePoint Cloud
🎯 What it does: This paper presents DroneSplat, a robust 3D Gaussian splatting framework for outdoor drone imagery, aimed at removing dynamic interference and achieving high-quality 3D reconstruction under sparse viewpoint conditions.
DropGaussian: Structural Regularization for Sparse-view Gaussian Splatting
Hyunwoo Park (Konkuk University), Wonjun Kim (Konkuk University)
Gaussian SplattingPoint CloudBenchmark
🎯 What it does: Proposes the DropGaussian method, which randomly drops some 3D Gaussians during sparse perspective training to alleviate overfitting.
DropoutGS: Dropping Out Gaussians for Better Sparse-view Rendering
Yexing Xu (Sun Yat-Sen University), Yulan Guo (Sun Yat-Sen University)
RestorationData SynthesisGaussian SplattingPoint Cloud
🎯 What it does: The DropoutGS method is proposed, which incorporates random Gaussian point dropout regularization and edge-based refinement strategies in 3D Gaussian point rendering to alleviate overfitting and detail loss issues under sparse viewpoints.
DrVideo: Document Retrieval Based Long Video Understanding
Ziyu Ma (Hunan University), Jianfei Cai (Monash University)
RetrievalTransformerLarge Language ModelVision Language ModelVideoTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A long video understanding framework based on document retrieval, DrVideo, is proposed, which transforms long videos into long text documents and utilizes LLM for question answering.
DSPNet: Dual-vision Scene Perception for Robust 3D Question Answering
Jingzhou Luo (Sun Yat-sen University), Liang Lin (Sun Yat-sen University)
RecognitionRetrievalTransformerVision Language ModelImageTextPoint Cloud
🎯 What it does: This paper proposes a Dual-Sight Perception Network (DSPNet) that integrates multi-view images and 3D point clouds to accomplish the 3D visual question answering (3D QA) task.
DSV-LFS: Unifying LLM-Driven Semantic Cues with Visual Features for Robust Few-Shot Segmentation
Amin Karimi (Concordia University), Charalambos Poullis (Concordia University)
SegmentationLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: The DSV-LFS framework is proposed, which combines semantic prompts generated by large language models with pixel-level matching for image support, used for few-shot semantic segmentation.
DTGBrepGen: A Novel B-rep Generative Model through Decoupling Topology and Geometry
Jing Li (University of Science and Technology of China), Falai Chen (University of Science and Technology of China)
GenerationTransformerDiffusion modelMesh
🎯 What it does: A B-rep generation framework that separates topology and geometry is proposed, first generating a valid topological structure and then gradually generating geometry;
DTOS: Dynamic Time Object Sensing with Large Multimodal Model
Jirui Tian (Dalian University of Technology), Gao Huang (Dalian University of Technology)
Object DetectionRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodality
🎯 What it does: A dynamic spatiotemporal target perception framework (DTOS) based on large language models is proposed to address the issues of multi-reference and visual information loss in RVOS and Moment Retrieval.
Dual Consolidation for Pre-Trained Model-Based Domain-Incremental Learning
Da-Wei Zhou (Nanjing University), De-Chuan Zhan (Nanjing University)
Domain AdaptationRepresentation LearningTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes a Dual Integration (DUCT) method to address the issues of representation drift and classifier mismatch in domain incremental learning with pre-trained models.
Dual Diffusion for Unified Image Generation and Understanding
Zijie Li (Carnegie Mellon University), Peng Wang (ByteDance)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes a dual-branch diffusion transformer that jointly learns the distribution of images and text, achieving multimodal tasks such as image generation, image-to-text generation, caption generation, and visual question answering.
Dual Energy-Based Model with Open-World Uncertainty Estimation for Out-of-distribution Detection
Qi Chen (University of Science and Technology of China), Hu Ding (University of Science and Technology of China)
ClassificationAnomaly DetectionImage
🎯 What it does: A new dual energy model (DEBO) is proposed for detecting out-of-distribution samples (OOD), addressing the limitations of existing methods through a dual classifier architecture and a unified energy-based objective function.
Dual Exposure Stereo for Extended Dynamic Range 3D Imaging
Juhyung Choi (POSTECH), Seung-Hwan Baek (POSTECH)
RestorationData SynthesisDepth EstimationAutonomous DrivingOptical FlowImageVideo
🎯 What it does: This paper proposes a Dual-Exposure Stereo method for achieving extended dynamic range 3D image reconstruction under varying lighting conditions.
Dual Focus-Attention Transformer for Robust Point Cloud Registration
Kexue Fu (Qilu University of Technology), Longxiang Gao
TransformerPoint Cloud
🎯 What it does: Proposes the Dual Focus-Attention Transformer (DFAT), which achieves point cloud registration by focusing on sparse key points at both coarse and fine scales.
Dual Prompting Image Restoration with Diffusion Transformers
Dehong Kong (Sun Yat-sen University), WenQi Ren (Chinese University of Hong Kong)
RestorationSuper ResolutionTransformerDiffusion modelRectified FlowImage
🎯 What it does: A dual-prompt image restoration method called DPIR has been designed and implemented, achieving high-quality image restoration in the Diffusion Transformer using dual conditions of vision and text.
Dual Semantic Guidance for Open Vocabulary Semantic Segmentation
Zhengyang Wang (Tianjin University), Liang Wan (Tianjin University)
SegmentationContrastive LearningImageText
🎯 What it does: A Dual Semantic Guidance framework is proposed, which extracts visual and textual semantics from image instances and text nouns to guide the CLIP model in achieving unsupervised open vocabulary semantic segmentation.
Dual-Agent Optimization framework for Cross-Domain Few-Shot Segmentation
Zhaoyang Li (University of Science and Technology of China), Xiang Liu (Dongguan University of Technology)
SegmentationDomain AdaptationOptimizationAdversarial AttackConvolutional Neural NetworkTransformerImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A Dual-Agent Optimization (DATO) framework is proposed to simultaneously address the issues of feature domain sensitivity and matching process sensitivity in cross-domain few-shot segmentation.
Dual-Granularity Semantic Guided Sparse Routing Diffusion Model for General Pansharpening
Yinghui Xing (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
Image TranslationRestorationTransformerMixture of ExpertsVision Language ModelDiffusion modelImage
🎯 What it does: This paper proposes a dual-granularity semantic-guided diffusion model SGDiff based on sparse expert routing for achieving general full-resolution image fusion.
Dual-Interrelated Diffusion Model for Few-Shot Anomaly Image Generation
Ying Jin (Fudan University), Yabiao Wang (Zhejiang University)
GenerationAnomaly DetectionDiffusion modelImage
🎯 What it does: Designed and implemented the DualAnoDiff dual-branch diffusion model, capable of simultaneously generating anomalous images and corresponding masks, achieving high-quality, aligned anomalous image-mask pairs.
Dual-view X-ray Detection: Can AI Detect Prohibited Items from Dual-view X-ray Images like Humans?
Renshuai Tao (Beijing Jiaotong University), Yao Zhao (Beijing Jiaotong University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: A large-scale dual-view X-ray image dataset, LDXray, is proposed, along with the Auxiliary-View Enhanced Network (AENet) dual-view detection framework, which enhances the localization accuracy of hard-to-detect categories by utilizing the cross-information between the main view and the auxiliary view, as well as expert models.
DualPM: Dual Posed-Canonical Point Maps for 3D Shape and Pose Reconstruction
Ben Kaye (University of Oxford), Andrea Vedaldi (University of Oxford)
Pose EstimationConvolutional Neural NetworkTransformerImagePoint Cloud
🎯 What it does: Predicting dual point mapping DualPM from a single image to achieve 3D shape and pose reconstruction of deformable objects.