CVPR 2023 Papers — Page 5
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2353 papers
Correspondence Transformers With Asymmetric Feature Learning and Matching Flow Super-Resolution
Yixuan Sun (Fudan University), Weifeng Ge (Fudan University)
Image TranslationSuper ResolutionTransformerImage
🎯 What it does: This paper proposes a novel Transformer-based semantic correspondence framework—ACTR, which achieves end-to-end training of dense visual correspondence through asymmetric feature learning and matching flow super-resolution.
COT: Unsupervised Domain Adaptation With Clustering and Optimal Transport
Yang Liu (Alibaba Group), Baigui Sun (Alibaba Group)
Domain AdaptationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a clustering-based optimal transport (COT) method for unsupervised domain adaptation, which enhances feature alignment by performing optimal transport between the cluster centers of the source and target domains.
CoWs on Pasture: Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation
Samir Yitzhak Gadre (Columbia University), Shuran Song (Columbia University)
Object DetectionRobotic IntelligenceTransformerLarge Language ModelVision Language ModelMultimodalityBenchmarkAgriculture Related
🎯 What it does: This study investigates language-driven zero-shot object navigation (L-ZSON), proposes the CLIP on Wheels (CoW) baseline, and constructs the PASTURE evaluation benchmark, systematically assessing 22 different combinations.
CP3: Channel Pruning Plug-In for Point-Based Networks
Yaomin Huang (East China Normal University), Jian Tang (Midea Group)
Object DetectionSegmentationConvolutional Neural NetworkPoint Cloud
🎯 What it does: A general channel pruning plugin CP3 is proposed to transfer the channel pruning method of 2D CNNs to 3D point cloud networks (PNN), enhancing the pruning effect through coordinate enhancement and knowledge recovery.
CR-FIQA: Face Image Quality Assessment by Learning Sample Relative Classifiability
Fadi Boutros (Fraunhofer Institute for Computer Graphics Research IGD), Naser Damer (Fraunhofer Institute for Computer Graphics Research IGD)
ClassificationRecognitionSupervised Fine-TuningImage
🎯 What it does: This paper proposes a new facial image quality assessment method CR-FIQA, which predicts image quality by utilizing the relative classifiability of samples during the training process.
CRAFT: Concept Recursive Activation FacTorization for Explainability
Thomas Fel (Brown University), Thomas Serre (Brown University)
Explainability and InterpretabilityImage
🎯 What it does: This paper proposes a concept-based explainable method called CRAFT, which automatically extracts interpretable high-level concepts by recursively decomposing network activations. It evaluates the importance of concepts using Sobol sensitivity analysis and implements concept attribution maps through implicit differentiation to explain the model's decision-making process on images.
CREPE: Can Vision-Language Foundation Models Reason Compositionally?
Zixian Ma (Stanford University), Ranjay Krishna (University of Washington)
RetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Designed and released the CREPE benchmark to evaluate the combinatorial reasoning capabilities of visual language models in terms of systematicity and productivity.
Critical Learning Periods for Multisensory Integration in Deep Networks
Michael Kleinman (University of California), Stefano Soatto (AWS AI Labs)
ClassificationTransformerImageVideo
🎯 What it does: The paper studies the critical learning period that exists in deep networks during multi-source information fusion and proves that the dependence on source information in the early training stage is crucial, while shallow networks do not exhibit this phenomenon.
CrOC: Cross-View Online Clustering for Dense Visual Representation Learning
Thomas Stegmüller (École Polytechnique Fédérale de Lausanne), Jean-Philippe Thiran (École Polytechnique Fédérale de Lausanne)
SegmentationRepresentation LearningTransformerContrastive LearningImageVideo
🎯 What it does: This paper proposes a cross-view online clustering method CrOC for dense visual representation learning of unlabeled scene images.
Cross-Domain 3D Hand Pose Estimation With Dual Modalities
Qiuxia Lin (National University of Singapore), Angela Yao (National University of Singapore)
Pose EstimationDomain AdaptationKnowledge DistillationConvolutional Neural NetworkSupervised Fine-TuningContrastive LearningImageMultimodality
🎯 What it does: A dual-modal network is proposed, utilizing synthetic data's RGB and depth maps for pre-training, and performing semi-supervised fine-tuning on unlabeled real RGB data to address the domain discrepancy problem in cross-domain gesture pose estimation.
Cross-Domain Image Captioning With Discriminative Finetuning
Roberto Dessì (Meta AI), Marco Baroni (UPF)
GenerationRetrievalDomain AdaptationTransformerReinforcement LearningVision Language ModelImageText
🎯 What it does: This paper proposes an unsupervised discriminative fine-tuning method that performs self-supervised fine-tuning on existing image captioning models, allowing the generated captions to assist a frozen image retriever in accurately identifying the target image among a set of candidate images, thereby improving cross-domain caption generation performance.
Cross-GAN Auditing: Unsupervised Identification of Attribute Level Similarities and Differences Between Pretrained Generative Models
Matthew L. Olson (Oregon State University), Weng-Keen Wong (Oregon State University)
GenerationData SynthesisDomain AdaptationGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: This paper proposes an unsupervised cross-GAN auditing framework called xGA, which is used to compare the semantic attributes between reference GANs and client GANs, automatically identifying three types of attributes: common attributes, client-specific attributes, and missing attributes.
Cross-Guided Optimization of Radiance Fields With Multi-View Image Super-Resolution for High-Resolution Novel View Synthesis
Youngho Yoon (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
Data SynthesisSuper ResolutionOptimizationNeural Radiance FieldImage
🎯 What it does: This paper proposes a Cross-guided Optimization Framework (CROP) that enhances the quality of High-Resolution New View Synthesis (HRNVS) by performing Multi-View Image Super-Resolution (MVSR) on training views during the NeRF optimization process.
Cross-Image-Attention for Conditional Embeddings in Deep Metric Learning
Dmytro Kotovenko (LMU Munich), Björn Ommer (LMU Munich)
RetrievalRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: During the training process, the encoding of the image to be embedded is combined with the embedding of another image through cross-image attention, thereby learning conditional embeddings, while unconditional embeddings are still used for retrieval during the inference phase.
Cross-Modal Implicit Relation Reasoning and Aligning for Text-to-Image Person Retrieval
Ding Jiang (Wuhan University), Mang Ye (Wuhan University)
RetrievalTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: This paper proposes a cross-modal implicit relationship reasoning and alignment framework (IRRA), which utilizes the Masked Language Modeling (MLM) task to achieve fine-grained relationship learning between vision and text, and optimizes cross-modal matching through Similarity Distribution Matching loss.
Crossing the Gap: Domain Generalization for Image Captioning
Yuchen Ren (University of Science and Technology of China), Wanli Ouyang (University of Science and Technology of China)
GenerationDomain AdaptationTransformerContrastive LearningImageTextBenchmark
🎯 What it does: This paper proposes the Domain Generalization for Image Captioning (DGIC) task and establishes a benchmark across five domains (MSCOCO, VizWiz, Flickr30k, CUB-200, Oxford-102); it also introduces the Language-guided Semantic Metric Learning (LSML) framework to enhance the model's generalization ability in unseen target domains.
Crowd3D: Towards Hundreds of People Reconstruction From a Single Image
Hao Wen (Tianjin University), Kun Li (Tianjin University)
Pose EstimationConvolutional Neural NetworkImageBenchmark
🎯 What it does: Reconstructing the 3D poses, shapes, and global positions of hundreds of people from a single large scene image to achieve a globally consistent 3D model of the crowd.
CrowdCLIP: Unsupervised Crowd Counting via Vision-Language Model
Dingkang Liang (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
Object DetectionSegmentationTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes CrowdCLIP, an unsupervised crowd counting method that utilizes the CLIP vision-language model and achieves this through ranking-based contrastive learning.
CUDA: Convolution-Based Unlearnable Datasets
Vinu Sankar Sadasivan (University of Maryland), Soheil Feizi (University of Maryland)
Safty and PrivacyData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: A new convolution-based unlearnable dataset generation technique (CUDA) is proposed, aimed at generating unlearnable datasets by controlling class convolutions to protect data privacy.
CUF: Continuous Upsampling Filters
Cristina N. Vasconcelos (Google Research), Andrea Tagliasacchi (Google Research)
RestorationSuper ResolutionNeural Radiance FieldImage
🎯 What it does: This paper proposes a continuous upsampling filter (CUF) based on neural fields for efficient single-image super-resolution.
Curricular Contrastive Regularization for Physics-Aware Single Image Dehazing
Yu Zheng (Ocean University of China), Yong Du (Ocean University of China)
RestorationConvolutional Neural NetworkTransformerContrastive LearningImagePhysics Related
🎯 What it does: This study focuses on single image dehazing and proposes a curriculum contrastive regularization and physics-aware dual-branch network called C2PNet.
Curricular Object Manipulation in LiDAR-Based Object Detection
Ziyue Zhu (Nankai University), Jian Yang (Nankai University)
Object DetectionAutonomous DrivingPoint Cloud
🎯 What it does: A framework that integrates curriculum learning into LiDAR 3D object detection is proposed, which includes two parts: loss reweighting and data augmentation.
Curvature-Balanced Feature Manifold Learning for Long-Tailed Classification
Yanbiao Ma (Xidian University), Lingling Li (Xidian University)
ClassificationImage
🎯 What it does: This paper studies the model bias caused by curvature imbalance in long-tail classification and proposes curvature regularization (CR) based on Gaussian curvature and dynamic curvature regularization (DCR) to achieve curvature balance and feature flattening.
Cut and Learn for Unsupervised Object Detection and Instance Segmentation
Xudong Wang (Meta AI), Ishan Misra (University of California Berkeley)
Object DetectionSegmentationTransformerContrastive LearningImageVideo
🎯 What it does: Using features extracted by the self-supervised visual Transformer (DINO), multiple rough object masks are generated through MaskCut, and then a general object detection and instance segmentation model is trained using DropLoss and multiple rounds of self-training. The resulting model can perform zero-shot detection and segmentation across various visual domains without using any annotations.
CutMIB: Boosting Light Field Super-Resolution via Multi-View Image Blending
Zeyu Xiao (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
RestorationSuper ResolutionConvolutional Neural NetworkTransformerImage
🎯 What it does: A data augmentation method named CutMIB, which stands for 'Crop-Mix-Paste', is designed and implemented. It enhances the training effect of light field super-resolution networks by randomly cropping low-resolution patches from the same position in multi-view light fields, averaging and mixing them, and then pasting them onto the corresponding high-resolution views, all while keeping the network structure unchanged.
CVT-SLR: Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment
Jiangbin Zheng (Westlake University), Stan Z. Li (Westlake University)
RecognitionConvolutional Neural NetworkRecurrent Neural NetworkAuto EncoderContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes a single-stream continuous sign language recognition framework called CVT-SLR based on contrastive visual-text transformation, which integrates pre-trained visual models with variational autoencoders to incorporate pre-trained linguistic knowledge and enhances performance through contrastive cross-modal alignment.
CXTrack: Improving 3D Point Cloud Tracking With Contextual Information
Tian-Xing Xu (Tsinghua University), Song-Hai Zhang (Tsinghua University)
Object TrackingAutonomous DrivingTransformerPoint Cloud
🎯 What it does: This paper proposes a Transformer-based 3D single object tracking framework called CXTrack, which fully utilizes contextual information between consecutive frames for tracking.
D2Former: Jointly Learning Hierarchical Detectors and Contextual Descriptors via Agent-Based Transformers
Jianfeng He (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
Object DetectionPose EstimationRetrievalTransformerAgentic AIImage
🎯 What it does: An end-to-end image matching framework named D Former is proposed, capable of jointly learning hierarchical detectors and contextual descriptors to achieve robust pixel-level matching.
DA Wand: Distortion-Aware Selection Using Neural Mesh Parameterization
Richard Liu, Rana Hanocka
SegmentationOptimizationGraph Neural NetworkMeshBenchmark
🎯 What it does: This paper predicts a local grid subset near user-selected points by training a neural network, ensuring low distortion and large area during UV mapping, and enabling quick results in interaction.
DA-DETR: Domain Adaptive Detection Transformer With Information Fusion
Jingyi Zhang (Nanyang Technological University), Shijian Lu (Nanyang Technological University)
Object DetectionDomain AdaptationTransformerImage
🎯 What it does: This paper proposes an unsupervised domain adaptation detection Transformer—DA-DETR, which utilizes an information fusion mechanism to achieve knowledge transfer between the source domain and the target domain.
DAA: A Delta Age AdaIN Operation for Age Estimation via Binary Code Transformer
Ping Chen (Jiayu Intelligent Technology Co., Ltd.), Zongjie Jiang (Jiayu Intelligent Technology Co., Ltd.)
RecognitionConvolutional Neural NetworkImage
🎯 What it does: A framework for age estimation based on Delta Age AdaIN (DAA) operation and binary code mapping is proposed, which transforms the features of the input image into 100 age difference features through AdaIN, and then uses AgeDecoder for regression.
DaFKD: Domain-Aware Federated Knowledge Distillation
Haozhao Wang (Huazhong University of Science and Technology), Zhigang Zeng (Soochow University)
Federated LearningKnowledge DistillationGenerative Adversarial NetworkImage
🎯 What it does: A domain-aware federated knowledge distillation method, DaFKD, is proposed, which uses a domain discriminator to weight the importance of local models from different clients on distilled samples, addressing the issue of data heterogeneity in federated learning.
DANI-Net: Uncalibrated Photometric Stereo by Differentiable Shadow Handling, Anisotropic Reflectance Modeling, and Neural Inverse Rendering
Zongrui Li (Nanyang Technological University), Xudong Jiang (Nanyang Technological University)
Depth EstimationImageBenchmark
🎯 What it does: A calibration-free photometric stereo method named DANI-Net is proposed, which estimates depth, surface normals, lighting direction/intensity, and spatially varying BRDF using differentiable shadow handling and an anisotropic reflection model in a joint inverse rendering framework.
DARE-GRAM: Unsupervised Domain Adaptation Regression by Aligning Inverse Gram Matrices
Ismail Nejjar (Ecole Polytechnique Federale de Lausanne), Olga Fink (Ecole Polytechnique Federale de Lausanne)
Domain AdaptationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an unsupervised domain adaptation regression method called DARE-GRAM based on aligned inverse Gram matrices, utilizing the closed-form OLS solution of linear regression to guide feature space alignment.
DART: Diversify-Aggregate-Repeat Training Improves Generalization of Neural Networks
Samyak Jain (Indian Institute of Science), R. Venkatesh Babu (Indian Institute of Science)
Domain AdaptationOptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper studies a diversified, aggregated, and repeated model weight strategy during training (DART) to enhance the generalization ability of neural networks both within and outside the domain.
DartBlur: Privacy Preservation With Detection Artifact Suppression
Baowei Jiang (Tsinghua University), Lu Fang (Tsinghua University)
Object DetectionSafty and PrivacyAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A learning-based fuzzification method called DartBlur is proposed, which suppresses detection artifacts generated during training while preserving privacy.
Data-Driven Feature Tracking for Event Cameras
Nico Messikommer (University of Zurich), Davide Scaramuzza (University of Zurich)
Object TrackingConvolutional Neural NetworkVideo
🎯 What it does: The first data-driven event camera feature tracker based on deep learning is proposed, achieving low-latency tracking by combining gray frame templates with event streams.
Data-Efficient Large Scale Place Recognition With Graded Similarity Supervision
María Leyva-Vallina (University of Groningen), Nicolai Petkov (University of Groningen)
RecognitionRetrievalConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes a hierarchical similarity for re-labeling visual localization datasets based on camera pose or 3D overlap information, and designs a Generalized Contrastive Loss (GCL) on this basis, achieving efficient training without hard negative sampling.
Data-Free Knowledge Distillation via Feature Exchange and Activation Region Constraint
Shikang Yu (Institute of Computing Technology Chinese Academy of Sciences), Shuqiang Jiang (Institute of Computing Technology Chinese Academy of Sciences)
GenerationKnowledge DistillationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A data-free knowledge distillation method called SpaceshipNet is proposed, which is based on Channel-level Feature Exchange (CFE) and Multi-Scale Spatial Activation Region Consistency (mSARC) constraints. It enables high-quality student model training by generating diverse synthetic images without accessing the original data.
Data-Free Sketch-Based Image Retrieval
Abhra Chaudhuri (University of Exeter), Anjan Dutta (University of Surrey)
RetrievalKnowledge DistillationContrastive LearningImage
🎯 What it does: A framework for sketch-based image retrieval (DF-SBIR) is proposed, which performs cross-modal metric space learning for photos and sketches using a pre-trained unimodal classifier without any training data.
DATE: Domain Adaptive Product Seeker for E-Commerce
Haoyuan Li (Zhejiang University), Zhou Zhao (Alibaba Group)
Object DetectionRetrievalDomain AdaptationTransformerContrastive LearningImageTextMultimodality
🎯 What it does: Construct a unified DATE framework to address product retrieval (PR) and localization (PG) tasks, and achieve unsupervised domain adaptation (PG-DA).
DATID-3D: Diversity-Preserved Domain Adaptation Using Text-to-Image Diffusion for 3D Generative Model
Gwanghyun Kim (Seoul National University), Se Young Chun (Seoul National University)
GenerationDomain AdaptationDiffusion modelImage
🎯 What it does: This paper develops a domain adaptation method for 3D generative models without additional target domain data, while preserving the diversity of text prompts.
DBARF: Deep Bundle-Adjusting Generalizable Neural Radiance Fields
Yu Chen (National University of Singapore), Gim Hee Lee (National University of Singapore)
Depth EstimationOptimizationRecurrent Neural NetworkNeural Radiance FieldImage
🎯 What it does: A Deep Bundle Adjustment Framework (DBARF) is proposed, which can jointly train a General Neural Radiance Field (GeNeRF) and a camera pose optimizer for self-supervised optimization of multi-view camera poses without requiring absolute camera pose initialization.
DC2: Dual-Camera Defocus Control by Learning To Refocus
Hadi Alzayer (Google), Abhishek Kar (Google)
RestorationDepth EstimationConvolutional Neural NetworkOptical FlowImage
🎯 What it does: A dual-camera-based post-depth control framework DC2 is proposed, which enables image refocusing, depth control, and depth-of-field effect synthesis.
DCFace: Synthetic Face Generation With Dual Condition Diffusion Model
Minchul Kim (Michigan State University), Xiaoming Liu (Michigan State University)
RecognitionGenerationData SynthesisDiffusion modelImage
🎯 What it does: Designed and implemented a dual-condition diffusion model DCFace to generate a diverse and label-consistent synthetic face dataset to enhance the training effectiveness of facial recognition models.
Dealing With Cross-Task Class Discrimination in Online Continual Learning
Yiduo Guo (Peking University), Dongyan Zhao (Peking University)
ClassificationRecognitionReinforcement LearningImage
🎯 What it does: A solution to the cross-task category discrimination (CTCD) problem in online continual learning, named GSA, is proposed.
DeAR: Debiasing Vision-Language Models With Additive Residuals
Ashish Seth (Indian Institute of Technology Madras), Chirag Agarwal (Adobe Inc)
ClassificationRecognitionData-Centric LearningTransformerSupervised Fine-TuningVision Language ModelImageVideoMultimodality
🎯 What it does: The DEAR framework is proposed to eliminate bias in visual language models by learning additive residuals, and a context-based PATA dataset is constructed for evaluation.
Decentralized Learning With Multi-Headed Distillation
Andrey Zhmoginov (Google AI), Max Vladymyrov (Google AI)
Federated LearningKnowledge DistillationImage
🎯 What it does: A decentralized learning framework based on multi-head distillation is proposed, allowing clients with private non-IID data to achieve knowledge exchange through public unlabeled data without sharing data, weights, or gradients.
DeCo: Decomposition and Reconstruction for Compositional Temporal Grounding via Coarse-To-Fine Contrastive Ranking
Lijin Yang (University of Tokyo), Norimasa Kobori (Woven by Toyota)
TransformerPrompt EngineeringContrastive LearningVideoText
🎯 What it does: A combinatorial temporal localization model DeCo based on coarse-to-fine decomposition and reorganization of query sentences and contrastive ranking is proposed, utilizing weakly supervised methods to learn the semantic correspondence between videos and clauses.
Decompose More and Aggregate Better: Two Closer Looks at Frequency Representation Learning for Human Motion Prediction
Xuehao Gao (Xi'an Jiaotong University), Yang Yang (Tencent AI Lab)
Pose EstimationRepresentation LearningGraph Neural NetworkVideo
🎯 What it does: This paper proposes a two-stage decomposition aggregation framework based on the frequency domain to improve the robustness of 3D human motion prediction.
Decompose, Adjust, Compose: Effective Normalization by Playing With Frequency for Domain Generalization
Sangrok Lee (Yonsei University), Ha Young Kim (Yonsei University)
Domain AdaptationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Proposed frequency domain-based normalization methods PCNorm, CCNorm, and SCNorm to improve domain generalization performance.
Decomposed Cross-Modal Distillation for RGB-Based Temporal Action Detection
Pilhyeon Lee (Yonsei University), Hyeran Byun (Yonsei University)
RecognitionKnowledge DistillationOptical FlowVideoMultimodality
🎯 What it does: By using a decomposed cross-modal distillation approach, RGB features and motion features are learned separately, and a temporal action detection model using only RGB is achieved through local attention fusion.
Decomposed Soft Prompt Guided Fusion Enhancing for Compositional Zero-Shot Learning
Xiaocheng Lu (Hong Kong Polytechnic University), Jingcai Guo (Hong Kong Polytechnic University)
RecognitionOptimizationTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: Addressing Combinatorial Zero-Shot Learning (CZSL) to achieve recognition and inference of unseen state-object combinations.
Decoupled Multimodal Distilling for Emotion Recognition
Yong Li (Nanjing University of Science and Technology), Zhen Cui (Nanjing University of Science and Technology)
RecognitionKnowledge DistillationTransformerMultimodality
🎯 What it does: A decoupled multimodal distillation framework (DMD) is proposed, which divides multimodal features into shared (modal-independent) and exclusive (modal-specific) parts, using homogeneous graph distillation and heterogeneous graph distillation to enhance emotion recognition performance.
Decoupled Semantic Prototypes Enable Learning From Diverse Annotation Types for Semi-Weakly Segmentation in Expert-Driven Domains
Simon Reiß (Karlsruhe Institute of Technology), Rainer Stiefelhagen (Karlsruhe Institute of Technology)
SegmentationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes a semi-weakly supervised semantic segmentation method called Decoupled Semantic Prototypes (DSP) that can simultaneously utilize pixel-level annotations, box annotations, point annotations, image-level labels, and unlabeled images.
Decoupling Human and Camera Motion From Videos in the Wild
Vickie Ye (University of California, Berkeley), Angjoo Kanazawa (University of California, Berkeley)
Object TrackingPose EstimationOptimizationSimultaneous Localization and MappingVideo
🎯 What it does: This paper proposes a method called SLAHMR, which can separate and recover the global trajectory of humans and camera motion in outdoor videos with free camera movement.
Decoupling Learning and Remembering: A Bilevel Memory Framework With Knowledge Projection for Task-Incremental Learning
Wenju Sun (Beijing Jiaotong University), Yangli-ao Geng (Beijing Jiaotong University)
ClassificationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: A dual-layer memory framework (BMKP) is proposed, which learns new tasks through working memory, stores knowledge in long-term memory, and compresses parameters in working memory into shared pattern bases using knowledge projection, which are then stored in long-term memory.
Decoupling MaxLogit for Out-of-Distribution Detection
Zihan Zhang (Huazhong University of Science and Technology), Xiang Xiang (Huazhong University of Science and Technology)
Anomaly DetectionConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes Decoupling MaxLogit (DML) and its improved version DML+, which splits MaxLogit into two parts: maximum cosine similarity (MaxCosine) and feature norm (MaxNorm). By employing methods such as cosine classifier, center loss, or focal loss during the training phase, it significantly enhances detection performance without external OOD data.
Decoupling-and-Aggregating for Image Exposure Correction
Yang Wang (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
RestorationConvolutional Neural NetworkImage
🎯 What it does: A Decoupling-and-Aggregating Convolution (DAConv) is proposed, which decouples contrast enhancement and detail recovery during the convolution process, and re-parameterizes the two branches into a single standard convolution during inference to improve exposure correction performance.
Deep Arbitrary-Scale Image Super-Resolution via Scale-Equivariance Pursuit
Xiaohang Wang (Shanghai Jiao Tong University), Yutian Liu (Shanghai Jiao Tong University)
RestorationSuper ResolutionTransformerImage
🎯 What it does: A scale-invariant arbitrary scale image super-resolution framework EQSR is proposed, which can achieve high-quality super-resolution for any scaling factor using the same model.
Deep Curvilinear Editing: Commutative and Nonlinear Image Manipulation for Pretrained Deep Generative Model
Takehiro Aoshima (Osaka University), Takashi Matsubara (Osaka University)
GenerationData SynthesisFlow-based ModelGenerative Adversarial NetworkImage
🎯 What it does: A method for latent space editing based on a curvilinear coordinate system, DeCurvEd, is proposed, and CurvilinearGANSpace is implemented on GANs to achieve nonlinear and interchangeable semantic attribute editing.
Deep Depth Estimation From Thermal Image
Ukcheol Shin (Korea Advanced Institute of Science and Technology), In So Kweon (Korea Advanced Institute of Science and Technology)
Depth EstimationTransformerImageMultimodality
🎯 What it does: This paper constructs a large-scale multispectral stereo dataset MS2, which includes RGB, NIR, thermal images, LiDAR, and GNSS/IMU information, and validates the performance of existing monocular and binocular depth estimation methods in the thermal image domain on this dataset.
Deep Deterministic Uncertainty: A New Simple Baseline
Jishnu Mukhoti (University of Oxford), Yarin Gal (University of Oxford)
ClassificationSegmentationAnomaly DetectionConvolutional Neural NetworkGaussian SplattingImage
🎯 What it does: A concise single forward inference model (DDU) is proposed, which achieves good regularization of the feature space by using spectral normalization on residual networks, and estimates the feature space density using Gaussian Discriminant Analysis (GDA) after training, thereby measuring the model's epistemic uncertainty and aleatoric uncertainty respectively.
Deep Discriminative Spatial and Temporal Network for Efficient Video Deblurring
Jinshan Pan (Nanjing University of Science and Technology), Jinhui Tang (China Electronics Technology Group Corporation)
RestorationConvolutional Neural NetworkOptical FlowVideo
🎯 What it does: This paper proposes DSTNet, a lightweight deep convolutional network for video deblurring, which is centered around a channel-gated dynamic network, a discriminative temporal feature fusion module, and a wavelet feature propagation mechanism, avoiding traditional optical flow or deformable convolution alignment methods.
Deep Dive Into Gradients: Better Optimization for 3D Object Detection With Gradient-Corrected IoU Supervision
Qi Ming (Beijing Institute of Technology), Yufei Guo (Peking University)
Object DetectionAutonomous DrivingOptimizationPoint CloudBenchmark
🎯 What it does: This paper studies and improves the IoU loss in 3D object detection, proposing the Gradient-Corrected IoU (GCIoU) loss and a gradient scaling strategy to achieve faster and more accurate bounding box regression.
Deep Factorized Metric Learning
Chengkun Wang (Tsinghua University), Jiwen Lu (Tsinghua University)
RetrievalTransformerImage
🎯 What it does: A deep factorized metric learning (DFML) framework is proposed, which dynamically allocates training samples to submodules of the Transformer model using learnable routers, achieving diversified feature extraction within the model.
Deep Fair Clustering via Maximizing and Minimizing Mutual Information: Theory, Algorithm and Metric
Pengxin Zeng (Sichuan University), Xi Peng (Sichuan University)
Auto EncoderTabularBiomedical Data
🎯 What it does: A deep fair clustering method FCMI based on mutual information theory is proposed, which unifies the maximization of conditional mutual information and the minimization of sensitive attribute mutual information to achieve compact, balanced, and fair clustering, and learns informative features through an autoencoder.
Deep Frequency Filtering for Domain Generalization
Shiqi Lin (University of Science and Technology of China), Zhibo Chen (University of Science and Technology of China)
ClassificationRetrievalDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: A Deep Frequency Filtering (DFF) module is proposed and implemented to dynamically and instance-adaptively enhance and suppress different frequency components of hidden layer features in the network, thereby improving the model's generalization ability on unseen domains.
Deep Graph Reprogramming
Yongcheng Jing (University of Sydney), Dacheng Tao (National University of Singapore)
Domain AdaptationKnowledge DistillationAdversarial AttackGraph Neural NetworkGraph
🎯 What it does: This paper proposes the Graph Adaptation and Reprogramming (GARE) framework, which utilizes a pre-trained GNN to accomplish various cross-domain and cross-task downstream learning without changing parameters or input features.
Deep Graph-Based Spatial Consistency for Robust Non-Rigid Point Cloud Registration
Zheng Qin (National University of Defense Technology), Kai Xu (National University of Defense Technology)
RecognitionOptimizationGraph Neural NetworkPoint Cloud
🎯 What it does: To address the issue of outlier removal in non-rigid point cloud registration, a graph-based spatial consistency network, GraphSCNet, is proposed to automatically filter inliers.
Deep Hashing With Minimal-Distance-Separated Hash Centers
Liangdao Wang (Sun Yat-Sen University), Ye Liu (Sun Yat-Sen University)
RetrievalOptimizationConvolutional Neural NetworkImage
🎯 What it does: Proposes a hash center optimization method based on the Gilbert-Varshamov bound constraint, and trains a deep hash network to achieve a theoretical guarantee of the minimum Hamming distance of the hash centers.
Deep Incomplete Multi-View Clustering With Cross-View Partial Sample and Prototype Alignment
Jiaqi Jin (National University of Defense Technology), En Zhu (National University of Defense Technology)
Representation LearningAuto EncoderImageBenchmark
🎯 What it does: A deep incomplete multi-view clustering framework named CPSPAN is proposed to learn consistent and distinguishable representations in the presence of missing samples.
Deep Learning of Partial Graph Matching via Differentiable Top-K
Runzhong Wang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
RecognitionObject DetectionSegmentationGraph Neural NetworkImageGraphBenchmark
🎯 What it does: This paper proposes a differentiable top-k selection module for achieving partial graph matching in deep graph matching networks, and presents two attention-based aggregation networks (AFA-I, AFA-U) for adaptive prediction of the number of matching inliers k. Additionally, it reconstructs and releases a visual graph matching benchmark for partial matching, IMC-PT-SparseGM, and provides a sparse implementation of NGMv2.
Deep Polarization Reconstruction With PDAVIS Events
Haiyang Mei (Dalian University of Technology), Tobi Delbruck (Institute of Neuroinformatics, University of Zurich and ETH Zurich)
RestorationConvolutional Neural NetworkRecurrent Neural NetworkVideo
🎯 What it does: Designed and implemented an end-to-end deep neural network E2P to directly reconstruct high dynamic range, low motion blur polarization intensity, polarization angle, and polarization degree from PDAVIS event streams.
Deep Random Projector: Accelerated Deep Image Prior
Taihui Li (University of Minnesota), Ju Sun (University of Minnesota)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: An improved deep image prior model called Deep Random Projector (DRP) is proposed, which significantly accelerates the optimization process of the original Deep Image Prior (DIP) while maintaining recovery quality by freezing randomly initialized network weights, reducing network depth, and incorporating explicit priors (TV).
Deep Semi-Supervised Metric Learning With Mixed Label Propagation
Furen Zhuang (University of Illinois at Urbana-Champaign), Pierre Moulin (University of Illinois at Urbana-Champaign)
RetrievalContrastive LearningImage
🎯 What it does: This paper proposes a semi-supervised metric learning framework that significantly improves the retrieval performance of CBIR tasks by identifying hard negative samples during the label propagation process and combining mixed label propagation.
Deep Stereo Video Inpainting
Zhiliang Wu (Nanjing University of Science and Technology), Yan Yan (Illinois Institute of Technology)
RestorationGenerationConvolutional Neural NetworkOptical FlowVideo
🎯 What it does: This paper proposes an end-to-end stereo video inpainting network called SVINet, which can fill in the missing areas of the left and right views while maintaining temporal coherence and disparity consistency.
DeepLSD: Line Segment Detection and Refinement With Deep Image Gradients
Rémi Pautrat (ETH Zurich), Marc Pollefeys (Microsoft)
Object DetectionSegmentationOptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper presents DeepLSD, a hybrid method that combines the line attraction field learned by deep networks with traditional handcrafted line segment detectors, achieving accurate and robust line segment detection in 'wild' images without real line annotations, and further improving accuracy through optimization.
DeepMAD: Mathematical Architecture Design for Deep Convolutional Neural Network
Xuan Shen (Alibaba Group), Yanzhi Wang (Northeastern University)
ClassificationRecognitionObject DetectionSegmentationConvolutional Neural NetworkImageVideo
🎯 What it does: Proposes DeepMAD - a mathematical framework based on information theory maximum entropy and network effectiveness constraints, which automatically generates high-performance CNN structures using constrained mathematical programming;
DeepMapping2: Self-Supervised Large-Scale LiDAR Map Optimization
Chao Chen (New York University), Chen Feng (New York University)
Autonomous DrivingOptimizationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: To address the problem of large-scale LiDAR point cloud mapping, the DeepMapping2 method is proposed, which improves the self-supervised optimization framework of the original DeepMapping, enabling high-precision global registration on long sequences containing thousands of frames.
DeepSolo: Let Transformer Decoder With Explicit Points Solo for Text Spotting
Maoyuan Ye (Wuhan University), Dacheng Tao (JD Explore Academy)
RecognitionObject DetectionTransformerImageText
🎯 What it does: This paper proposes DeepSolo, an end-to-end unified framework for text detection and recognition based on a Transformer single decoder.
DeepVecFont-v2: Exploiting Transformers To Synthesize Vector Fonts With Higher Quality
Yuqing Wang (Peking University), Zhouhui Lian (Peking University)
GenerationData SynthesisTransformerImage
🎯 What it does: A dual-modal font synthesis framework called DeepVecFont-v2 based on the Transformer structure is proposed, which directly generates high-quality vector fonts in an end-to-end manner.
DeFeeNet: Consecutive 3D Human Motion Prediction With Deviation Feedback
Xiaoning Sun (Nanjing University of Science and Technology), Jianfeng Lu (Tianjin AiForward Science and Technology)
Pose EstimationRecurrent Neural NetworkVideo
🎯 What it does: This paper proposes a lightweight module, DeFeeNet, which can be integrated into existing 3D human motion prediction models to achieve error awareness and feedback during continuous prediction processes, significantly improving accuracy in multi-round predictions.
Defending Against Patch-Based Backdoor Attacks on Self-Supervised Learning
Ajinkya Tejankar (University of California), Liang Tan (Meta AI)
Representation LearningAdversarial AttackTransformerContrastive LearningImage
🎯 What it does: Proposes the PatchSearch method for detecting and removing patch-based backdoor attack samples in self-supervised learning.
Defining and Quantifying the Emergence of Sparse Concepts in DNNs
Jie Ren (Shanghai Jiao Tong University), Quanshi Zhang (Shanghai Jiao Tong University)
OptimizationExplainability and InterpretabilityAdversarial AttackImageTextTabular
🎯 What it does: This paper proposes and proves that in sufficiently trained deep neural networks, reasoning logic can be represented by sparse interactive concepts (interaction patterns) and constructs them as causal graphs and And-Or graphs.
Deformable Mesh Transformer for 3D Human Mesh Recovery
Yusuke Yoshiyasu (National Institute of Advanced Industrial Science and Technology)
GenerationPose EstimationTransformerImageMesh
🎯 What it does: A decoder-level Transformer named DeFormer is designed to generate high-resolution 3D human meshes from a single RGB image, achieving efficient inference through multi-scale feature fusion and sparse self-attention/deformable cross-attention.
DegAE: A New Pretraining Paradigm for Low-Level Vision
Yihao Liu (Shanghai Artificial Intelligence Laboratory), Chao Dong (ShenZhen Key Lab of Computer Vision and Pattern Recognition)
RestorationConvolutional Neural NetworkTransformerAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: A low-level vision self-supervised pre-training framework called DegAE is proposed, which utilizes a degradation autoencoder to learn general image representations and fine-tunes on various low-level vision tasks.
DeGPR: Deep Guided Posterior Regularization for Multi-Class Cell Detection and Counting
Aayush Kumar Tyagi (Indian Institute of Technology Delhi), Mausam (Indian Institute of Technology Delhi)
Object DetectionContrastive LearningImageBiomedical Data
🎯 What it does: This paper proposes a deep-guided posterior regularization framework (DEGPR) for multi-class cell detection and counting.
DejaVu: Conditional Regenerative Learning To Enhance Dense Prediction
Shubhankar Borse (Qualcomm AI Research), Fatih Porikli (Qualcomm AI Research)
SegmentationDepth EstimationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: The DejaVu framework is proposed, which redacts input images during training and uses a conditional reconstruction module to restore the original image, thereby providing additional supervision for dense prediction models.
Delivering Arbitrary-Modal Semantic Segmentation
Jiaming Zhang (Karlsruhe Institute of Technology), Rainer Stiefelhagen (Karlsruhe Institute of Technology)
SegmentationAutonomous DrivingTransformerImageMultimodality
🎯 What it does: A semantic segmentation framework called CMNeXt is proposed, which can handle an arbitrary number of sensors, and a multimodal dataset named DELIVER is constructed.
Delving Into Discrete Normalizing Flows on SO(3) Manifold for Probabilistic Rotation Modeling
Yulin Liu (Peking University), He Wang (Peking University)
Pose EstimationFlow-based ModelImage
🎯 What it does: A discrete regularization flow for the SO(3) rotation manifold is proposed, utilizing Mobius coupling layers and quaternion affine transformations to construct invertible mappings that can map from simple benchmark distributions to any complex rotation distribution.
Delving Into Shape-Aware Zero-Shot Semantic Segmentation
Xinyu Liu (Xidian University), Guyue Zhou (Tsinghua University)
SegmentationTransformerContrastive LearningImage
🎯 What it does: This paper proposes a shape-aware zero-shot semantic segmentation framework SAZS, which can accurately segment unseen categories in the training set without retraining.
Delving StyleGAN Inversion for Image Editing: A Foundation Latent Space Viewpoint
Hongyu Liu (Hong Kong University of Science and Technology), Qifeng Chen (Hong Kong University of Science and Technology)
GenerationData SynthesisGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: For the real image inversion and editing of StyleGAN, the CLCAE method is proposed, which first aligns the image space and the basic latent space W using contrastive learning, and then transforms W into W+ and F through a cross-attention encoder to achieve high-quality reconstruction and editability.
Demystifying Causal Features on Adversarial Examples and Causal Inoculation for Robust Network by Adversarial Instrumental Variable Regression
Junho Kim (Korea Advanced Institute of Science and Technology), Yong Man Ro (Korea Advanced Institute of Science and Technology)
OptimizationAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: Using instrumental variable regression and maximum moment restrictions in adversarial zero-sum games, this paper identifies and extracts causal features from adversarial samples and injects them into defense networks to enhance robustness.
Dense Distinct Query for End-to-End Object Detection
Shilong Zhang (Shanghai AI Laboratory), Kai Chen (Shanghai AI Laboratory)
Object DetectionConvolutional Neural NetworkTransformerImage
🎯 What it does: The Dense Distinct Queries (DDQ) framework is proposed, which integrates dense and distinct queries to achieve end-to-end object detection, significantly improving the accuracy of models such as FCN, R‑CNN, and DETR.
Dense Network Expansion for Class Incremental Learning
Zhiyuan Hu (University of California San Diego), Nuno Vasconcelos (University of California San Diego)
ClassificationKnowledge DistillationTransformerImage
🎯 What it does: A Dense Network Expansion (DNE) framework is proposed to address the issues of catastrophic forgetting and rapid model growth in Class Incremental Learning (CIL).
Dense-Localizing Audio-Visual Events in Untrimmed Videos: A Large-Scale Benchmark and Baseline
Tiantian Geng (Southern University of Science and Technology), Feng Zheng (University of Birmingham)
RecognitionObject DetectionTransformerOptical FlowVideoMultimodalityBenchmarkAudio
🎯 What it does: This study proposes a dense localization method for audio-visual events in unedited videos and constructs the UnAV-100 dataset.
Density-Insensitive Unsupervised Domain Adaption on 3D Object Detection
Qianjiang Hu (Peking University), Wei Hu (Peking University)
Object DetectionDomain AdaptationAutonomous DrivingKnowledge DistillationGraph Neural NetworkPoint Cloud
🎯 What it does: This paper proposes a density-insensitive teacher-student framework for unsupervised domain adaptation in 3D object detection;
DepGraph: Towards Any Structural Pruning
Gongfan Fang (National University of Singapore), Xinchao Wang (National University of Singapore)
CompressionOptimizationConvolutional Neural NetworkRecurrent Neural NetworkGraph Neural NetworkTransformerImageTextPoint CloudGraph
🎯 What it does: This paper proposes a general 'Dependency Graph' (DepGraph) method for automatic structural pruning of any network architecture, capable of identifying and simultaneously pruning coupled parameter groups in one go.
Depth Estimation From Camera Image and mmWave Radar Point Cloud
Akash Deep Singh (University of California), Alex Wong (Yale University)
Depth EstimationAutonomous DrivingConvolutional Neural NetworkImagePoint Cloud
🎯 What it does: This paper proposes a depth estimation method based on single-frame camera images and millimeter-wave radar point clouds, employing a two-stage network: RadarNet first maps radar points to possible surfaces in the image, generating a semi-dense radar depth map; FusionNet then adaptively fuses radar depth with image features through a gating mechanism, ultimately outputting a dense depth map.