ICCV 2023 Papers — Page 4
IEEE/CVF International Conference on Computer Vision · 2156 papers
COMPASS: High-Efficiency Deep Image Compression with Arbitrary-scale Spatial Scalability
Jongmin Park (Korea Advanced Institute of Science and Technology), Munchurl Kim (Korea Advanced Institute of Science and Technology)
CompressionImage
🎯 What it does: A scale-agnostic, space-scalable image compression method based on deep learning called COMPASS is proposed;
Compatibility of Fundamental Matrices for Complete Viewing Graphs
Martin Bråtelund (University of Oslo), Felix Rydell (KTH Royal Institute of Technology)
🎯 What it does: This paper studies the algebraic conditions for determining whether a set of fundamental matrices is compatible under a complete view graph (i.e., with all view pairs' fundamental matrices known) and provides explicit homogeneous polynomial constraints. It further proves that in the case of four views or higher dimensions, compatibility of four views guarantees global compatibility. It also demonstrates that the eigenvalue conditions proposed by previous researchers regarding n-view fundamental matrices can be disregarded in general and collinear cases. Additionally, a general 'cyclic theorem' is proposed, applicable to any view graph.
Complementary Domain Adaptation and Generalization for Unsupervised Continual Domain Shift Learning
Wonguk Cho (Seoul National University), Taesup Kim (Seoul National University)
Domain AdaptationImage
🎯 What it does: A framework named CoDAG is proposed, which simultaneously achieves domain adaptation (DA) and domain generalization (DG) in unsupervised continuous domain shift learning through a complementary approach of two models, and implements knowledge sharing and forgetting suppression through techniques such as pseudo-labeling, noise robustness, and replay buffer.
Compositional Feature Augmentation for Unbiased Scene Graph Generation
Lin Li (Zhejiang University), Long Chen (Hong Kong University of Science and Technology)
Object DetectionGenerationTransformerContrastive LearningGraph
🎯 What it does: A bias-free scene graph generation method called CFA is designed, which increases the feature diversity of tail class relationship triples by utilizing intrinsic feature replacement and extrinsic feature mixing.
Computation and Data Efficient Backdoor Attacks
Yutong Wu (Nanyang Technological University), Tianwei Zhang (Nanyang Technological University)
Computational EfficiencyAdversarial AttackImagePoint Cloud
🎯 What it does: A fast and scalable backdoor sample selection method based on Representative Distance (RD) scores is proposed, which reduces the proportion of poisoned samples required for backdoor injection and improves the attack success rate.
Computational 3D Imaging with Position Sensors
Jeremy Klotz (Columbia University), Aswin C. Sankaranarayanan (Carnegie Mellon University)
Point Cloud
🎯 What it does: 3D scanning based on Position Sensitive Diodes (PSD) is implemented to solve the global illumination error problem.
Computationally-Efficient Neural Image Compression with Shallow Decoders
Yibo Yang (University of California), Stephan Mandt (University of California)
CompressionComputational EfficiencyImage
🎯 What it does: This paper proposes the use of shallow (even linear) decoding transformations in neural image compression, significantly reducing decoding complexity.
Concept-wise Fine-tuning Matters in Preventing Negative Transfer
Yunqiao Yang (City University of Hong Kong), Ying Wei (Tencent AI Lab)
ClassificationSegmentationDomain AdaptationConvolutional Neural NetworkTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: Proposes the Concept-Tuning approach, which fine-tunes pre-trained models at the level of concepts (patches) to address rare features and spurious correlations, significantly reducing negative transfer.
Conceptual and Hierarchical Latent Space Decomposition for Face Editing
Savas Ozkan (Samsung Research), Tom Robinson (Samsung Research)
GenerationData SynthesisPose EstimationTransformerGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a self-supervised encoder-decoder model based on Transformer, which decomposes the multi-resolution feature space of the pre-trained StyleGAN into a concept-level intermediate latent space. This space is used to train a GAN space controller to achieve precise editing of facial poses, expressions, and lighting while maintaining high fidelity and identity consistency.
Conditional 360-degree Image Synthesis for Immersive Indoor Scene Decoration
Ka Chun Shum (Hong Kong University of Science and Technology), Sai-Kit Yeung (Hong Kong University of Science and Technology)
Image TranslationGenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: An automatic decoration method is proposed for 360° panoramic indoor scenes: given a panoramic image with an empty background, generate a panoramic image with furniture.
Conditional Cross Attention Network for Multi-Space Embedding without Entanglement in Only a SINGLE Network
Chull Hwan Song (Dealicious Inc.), Yeong Hyeon Gu (Sejong University)
TransformerImage
🎯 What it does: A Conditional Cross-Attention Network (CCA) is proposed, which achieves multi-attribute space embedding in a single network by adding conditional encoding at the end of the Vision Transformer.
Confidence-aware Pseudo-label Learning for Weakly Supervised Visual Grounding
Yang Liu (Peking University), Yuxin Peng (Peking University)
Object DetectionRetrievalTransformerPrompt EngineeringVision Language ModelImageText
🎯 What it does: A confidence-aware pseudo-label learning framework for weakly supervised visual localization (CPL) is proposed, which constructs reliable image-text correspondences through the automatic generation of diverse pseudo queries, unified modality query propagation, and cross-modal confidence verification.
Confidence-based Visual Dispersal for Few-shot Unsupervised Domain Adaptation
Yizhe Xiong (Tsinghua University), Guiguang Ding (Tsinghua University)
Domain AdaptationImageBenchmark
🎯 What it does: A confidence-based visual dispersion (C-VisDiT) method is proposed for few-shot unsupervised domain adaptation (FUDA), enhancing knowledge transfer through cross-domain and same-domain visual dispersion.
Consistent Depth Prediction for Transparent Object Reconstruction from RGB-D Camera
Yuxiang Cai (Nankai University), Bo Ren (Nankai University)
Depth EstimationSimultaneous Localization and MappingVideoPoint Cloud
🎯 What it does: For RGB-D reconstruction of transparent objects, a depth prediction network based on plane sweeping stereo matching is proposed and embedded into ElasticFusion for real-time reconstruction.
ConSlide: Asynchronous Hierarchical Interaction Transformer with Breakup-Reorganize Rehearsal for Continual Whole Slide Image Analysis
Yanyan Huang (University of Hong Kong), Lequan Yu (University of Hong Kong)
ClassificationTransformerContrastive LearningImage
🎯 What it does: The ConSlide framework is proposed to achieve Whole Slide Image (WSI) analysis in a continuous learning environment, addressing challenges such as large image sizes, hierarchical structure utilization, and catastrophic forgetting.
Constraining Depth Map Geometry for Multi-View Stereo: A Dual-Depth Approach with Saddle-shaped Depth Cells
Xinyi Ye (Huazhong University of Science and Technology), Xin Li (University of Albany)
Depth EstimationConvolutional Neural NetworkPoint CloudBenchmark
🎯 What it does: A dual depth prediction and checkerboard selection strategy is proposed to enhance the accuracy of 3D reconstruction by constructing hanging-shaped depth cells in multi-view stereo.
ContactGen: Generative Contact Modeling for Grasp Generation
Shaowei Liu (University of Illinois Urbana-Champaign), Shenlong Wang (University of Illinois Urbana-Champaign)
GenerationRobotic IntelligencePoint Cloud
🎯 What it does: An object-based contact representation called ContactGen is proposed, which predicts contact information through a sequential CVAE; the predicted ContactGen is transformed into diverse, physically feasible hand grasp poses using a block-based Signed Distance Field hand model and an optimization solver.
Contactless Pulse Estimation Leveraging Pseudo Labels and Self-Supervision
Zhihua Li (Binghamton University), Lijun Yin (Binghamton University)
RecognitionOptimizationContrastive LearningVideo
🎯 What it does: A completely unsupervised remote photoplethysmography (rPPG) method is proposed, which combines self-supervised contrastive learning with traditional unsupervised pseudo-labeling to estimate pulse signals.
Content-Aware Local GAN for Photo-Realistic Super-Resolution
JoonKyu Park (Seoul National University), Kyoung Mu Lee (Seoul National University)
GenerationSuper ResolutionMixture of ExpertsGenerative Adversarial NetworkImage
🎯 What it does: Proposes the Content-Aware Local GAN (CAL-GAN) framework, which uses multiple expert discriminators to classify local features of different contents, thereby enhancing the realism of single-image super-resolution.
Context-Aware Planning and Environment-Aware Memory for Instruction Following Embodied Agents
Byeonghwi Kim (Yonsei University), Jonghyun Choi (Yonsei University)
Robotic IntelligenceRecurrent Neural NetworkReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes the CAPEAM framework, which enhances navigation and object interaction performance in interactive instruction following tasks through context-aware planning and environment-aware memory.
Continual Learning for Personalized Co-speech Gesture Generation
Chaitanya Ahuja (Carnegie Mellon University), Louis-Philippe Morency (Carnegie Mellon University)
GenerationData SynthesisDiffusion modelGenerative Adversarial NetworkMultimodality
🎯 What it does: A continuous learning-based co-speech gesture generation framework C-DiffGAN is proposed, enabling the model to quickly adapt to the gesture styles of new speakers while retaining the gesture styles of previous speakers with only 2-10 minutes of low-resource data for each speaker.
Continual Segment: Towards a Single, Unified and Non-forgetting Continual Segmentation Model of 143 Whole-body Organs in CT Scans
Zhanghexuan Ji (Alibaba Group), Dakai Jin (Alibaba Group)
SegmentationAnomaly DetectionKnowledge DistillationNeural Architecture SearchConvolutional Neural NetworkImageBiomedical DataComputed Tomography
🎯 What it does: Construct a sustainable and scalable multi-organ semantic segmentation model that supports continuous learning of new organ categories without accessing old data.
Continual Zero-Shot Learning through Semantically Guided Generative Random Walks
Wenxuan Zhang (King Abdullah University of Science and Technology), Mohamed Elhoseiny (King Abdullah University of Science and Technology)
ClassificationRecognitionGenerationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A continuous zero-shot learning framework (ICGZSL) is proposed that completely relies on unseen category semantic information, achieving realistic modeling of unseen visual spaces through generative models and random walk loss.
Continuously Masked Transformer for Image Inpainting
Keunsoo Ko (Catholic University of Korea), Chang-Su Kim (Korea University)
RestorationTransformerImage
🎯 What it does: This paper proposes a Continuous Mask-aware Transformer (CMT) for image inpainting, utilizing continuous masks and overlapping tokens for multi-layer masked self-attention and mask updates. After generating an initial inpainting result, a refinement network is employed to enhance details.
Contrastive Continuity on Augmentation Stability Rehearsal for Continual Self-Supervised Learning
Haoyang Cheng (University of Electronic Science and Technology of China), Hongliang Li (University of Electronic Science and Technology of China)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: To address the problem of catastrophic forgetting in Continuous Self-Supervised Learning (CSSL), two techniques are proposed: first, the design of 'Augmentation Stability Rehearsal (ASR)' selects the most representative and distinguishable samples for experience replay by calculating the augmentation stability of the samples; second, based on this, 'Contrastive Continuity on Augmentation Stability Rehearsal (C2ASR)' is proposed, which imposes a contrastive continuity constraint between replay samples and current task samples, combined with the Information Bottleneck principle, to achieve the retention of shared information and the elimination of redundant information.
Contrastive Feature Masking Open-Vocabulary Vision Transformer
Dahun Kim (Google DeepMind), Weicheng Kuo (Google DeepMind)
Object DetectionRetrievalTransformerContrastive LearningImage
🎯 What it does: This paper proposes a Contrastive Feature Masking Vision Transformer (CFM-ViT) for pre-training visual transformers to obtain richer pixel/region information, thereby enhancing open vocabulary object detection performance.
Contrastive Learning Relies More on Spatial Inductive Bias Than Supervised Learning: An Empirical Study
Yuanyi Zhong (University of Illinois at Urbana-Champaign), Yu-Xiong Wang (University of Illinois at Urbana-Champaign)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: By globally and locally shuffling the training set, the system evaluates the degree of dependence on spatial inductive bias in contrastive learning (CL) compared to supervised learning (SL).
Contrastive Model Adaptation for Cross-Condition Robustness in Semantic Segmentation
David Brüggemann (ETH Zurich), Luc Van Gool (ETH Zurich)
SegmentationDomain AdaptationContrastive LearningImage
🎯 What it does: Achieve adaptive cross-condition semantic segmentation models through contrastive learning without accessing labeled data from the source domain;
Contrastive Pseudo Learning for Open-World DeepFake Attribution
Zhimin Sun (Shanghai Jiao Tong University), Lizhuang Ma (Shanghai Jiao Tong University)
ClassificationRecognitionConvolutional Neural NetworkGenerative Adversarial NetworkContrastive LearningImageVideoBenchmark
🎯 What it does: A new Open World DeepFake Attribution (OW-DFA) benchmark is proposed, along with a Contrastive Pseudo Learning (CPL) framework designed to simultaneously identify known and unknown types of forgery in an open environment where labeled and unlabeled samples coexist.
Controllable Guide-Space for Generalizable Face Forgery Detection
Ying Guo (Meituan), Pengfei Yan (Meituan)
ClassificationDomain AdaptationConvolutional Neural NetworkContrastive LearningVideo
🎯 What it does: A controllable guide-space framework is proposed, combined with a decoupling module and a decision boundary adjustment module, to enhance the generalization ability of facial forgery detection models in unknown domains.
Controllable Person Image Synthesis with Pose-Constrained Latent Diffusion
Xiao Han (University of Surrey), Tao Xiang (University of Surrey)
GenerationData SynthesisPose EstimationComputational EfficiencyDiffusion modelImage
🎯 What it does: A latent diffusion model called PoCoLD based on DensePose is proposed for controllable human image synthesis.
Controllable Visual-Tactile Synthesis
Ruihan Gao, Jun-Yan Zhu (Carnegie Mellon University)
GenerationData SynthesisGenerative Adversarial NetworkImageMultimodality
🎯 What it does: This paper proposes a controllable visual-tactile synthesis method based on conditional generative adversarial networks, which can generate high-resolution visual images and corresponding tactile gradient information in one go based on user sketches (or text), and achieve a visually and haptically immersive experience through the TanvasTouch haptic screen.
Convex Decomposition of Indoor Scenes
Vaibhav Vavilala (University of Illinois at Urbana-Champaign), David Forsyth (University of Illinois at Urbana-Champaign)
SegmentationDepth EstimationOptimizationConvolutional Neural NetworkImage
🎯 What it does: Using deep learning to first predict the convex hull set of indoor scenes, and then refining the convex hull through gradient descent and backward selection, so that the final geometric decomposition can cover the entire scene and can be evaluated using standard depth, normal, and segmentation errors.
Convolutional Networks with Oriented 1D Kernels
Alexandre Kirchmeyer (Carnegie Mellon University), Jia Deng (Princeton University)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A fully 1D convolutional network is proposed, which achieves performance comparable to or even better than 2D convolution by introducing directional 1D convolution to replace traditional 2D convolution.
COOL-CHIC: Coordinate-based Low Complexity Hierarchical Image Codec
Théo Ladune (Orange Innovation), Thomas Leguay (Orange Innovation)
CompressionAuto EncoderImage
🎯 What it does: A low-complexity coordinate-based hierarchical image encoder called COOL-CHIC is proposed, which utilizes coordinate neural representation and multi-resolution latent variables for adaptive compression of a single image.
COOP: Decoupling and Coupling of Whole-Body Grasping Pose Generation
Yanzhao Zheng (Alibaba DAMO Academy), Wei Zhou (Alibaba DAMO Academy)
GenerationPose EstimationRobotic IntelligenceTransformerAuto EncoderPoint Cloud
🎯 What it does: Proposes the COOP framework, which can generate realistic full-body grasping poses for 3D target objects at different locations;
Coordinate Quantized Neural Implicit Representations for Multi-view Reconstruction
Sijia Jiang (Wayne State University), Zhizhong Han (Wayne State University)
GenerationData SynthesisOptimizationNeural Radiance FieldPoint CloudMesh
🎯 What it does: By quantizing continuous three-dimensional coordinates into discrete coordinates with extremely high resolution, mapping sampling points to discrete coordinates using nearest neighbor interpolation, and combining high-frequency positional encoding, we use volume rendering to learn neural implicit surfaces (occupancy fields or signed distance fields) under multi-view images.
Coordinate Transformer: Achieving Single-stage Multi-person Mesh Recovery from Videos
Haoyuan Li (Sun Yat-sen University), Xiaodan Liang (Sun Yat-sen University)
Object DetectionObject TrackingPose EstimationTransformerVideoMesh
🎯 What it does: A single-stage multi-person human mesh recovery framework (CoordFormer) is proposed, which can directly detect, track, and recover 3D human meshes of multiple people from videos simultaneously.
COPILOT: Human-Environment Collision Prediction and Localization from Egocentric Videos
Boxiao Pan (Stanford University), Leonidas J. Guibas (Stanford University)
Autonomous DrivingTransformerVideo
🎯 What it does: We propose a system called COPILOT that predicts and locates human-environment collisions from multi-view first-person videos, providing collision probabilities, involved joints, and collision heatmaps.
CopyRNeRF: Protecting the CopyRight of Neural Radiance Fields
Ziyuan Luo (Hong Kong Baptist University), Renjie Wan (Hong Kong Baptist University)
GenerationData SynthesisSafty and PrivacyNeural Radiance FieldImage
🎯 What it does: The CopyRNeRF framework is proposed, which protects the copyright of NeRF models by embedding watermarks in the NeRF color representation while maintaining invisibility and high quality during rendering;
CORE: Co-planarity Regularized Monocular Geometry Estimation with Weak Supervision
Yuguang Li (Samsung Research and Development Institute China), Feng Zhu (Samsung Research and Development Institute China)
Depth EstimationAutonomous DrivingConvolutional Neural NetworkImagePoint Cloud
🎯 What it does: A weakly supervised 3D geometry estimation method based on pixel-level depth-normal orthogonal constraints (CORE) is proposed, and a structure-aware normal estimator (SANE) is designed to achieve joint learning of depth and normals.
CORE: Cooperative Reconstruction for Multi-Agent Perception
Binglu Wang (Northwestern Polytechnical University), Tianfei Zhou (Beijing Institute of Technology)
Object DetectionSegmentationAutonomous DrivingConvolutional Neural NetworkPoint Cloud
🎯 What it does: This paper proposes and implements the CORE framework, which enhances the perception performance of multi-agent systems through collaborative reconstruction learning.
Corrupting Neuron Explanations of Deep Visual Features
Divyansh Srivastava (University of California San Diego), Tsui-Wei Weng (University of California San Diego)
Explainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper conducts a robustness analysis of neuron explanation methods (NEMs) and demonstrates that they can be significantly compromised when detecting data contaminated by random noise or carefully designed perturbations.
CoSign: Exploring Co-occurrence Signals in Skeleton-based Continuous Sign Language Recognition
Peiqi Jiao (Chinese Academy of Sciences), Xilin Chen (Xiaomi Inc.)
RecognitionPose EstimationGraph Neural NetworkSupervised Fine-TuningVideoMultimodality
🎯 What it does: This paper presents CoSign, a skeleton-based continuous sign language recognition method.
CoTDet: Affordance Knowledge Prompting for Task Driven Object Detection
Jiajin Tang (ShanghaiTech University), Sibei Yang (ShanghaiTech University)
Object DetectionSegmentationTransformerLarge Language ModelPrompt EngineeringImageChain-of-Thought
🎯 What it does: This paper proposes a multi-layer chain-of-thought prompting based on large language models to extract visual affordance knowledge and constructs the CoTDet framework, enabling the detector to utilize this knowledge to generate queries and guide box regression in task-driven object detection and instance segmentation.
Counterfactual-based Saliency Map: Towards Visual Contrastive Explanations for Neural Networks
Xue Wang (Zhejiang University), Kui Ren (Zhejiang University)
Explainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A unified visual contrast explanation framework CCE is proposed, which generates contrastive samples using adversarial sparse perturbations and explains three types of causal questions: 'Why P?', 'Why P instead of Q?', and 'Object a is P while object b is Q?'
Counting Crowds in Bad Weather
Zhi-Kai Huang (National Taiwan University), Ming-Hsuan Yang (UC Merced)
Object DetectionTransformerContrastive LearningImage
🎯 What it does: Developed AWCC-Net, achieving robust crowd counting under adverse weather conditions.
CPCM: Contextual Point Cloud Modeling for Weakly-supervised Point Cloud Semantic Segmentation
Lizhao Liu (South China University of Technology), Mingkui Tan (South China University of Technology)
SegmentationConvolutional Neural NetworkPoint Cloud
🎯 What it does: Research on weakly labeled point cloud semantic segmentation, proposing the CPCM model;
Creative Birds: Self-Supervised Single-View 3D Style Transfer
Renke Wang (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)
Image TranslationGenerationImage
🎯 What it does: A single-view 3D style transfer method is proposed, which can transfer the geometry and texture style of the target bird image to the source bird image, generating a new 3D bird model.
CRN: Camera Radar Net for Accurate, Robust, Efficient 3D Perception
Youngseok Kim (Korea Advanced Institute of Science and Technology), Dongsuk Kum (Korea Advanced Institute of Science and Technology)
Object DetectionObject TrackingSegmentationAutonomous DrivingComputational EfficiencyConvolutional Neural NetworkSimultaneous Localization and MappingMultimodalityPoint Cloud
🎯 What it does: This paper proposes Camera Radar Net (CRN), a 3D perception framework that utilizes the fusion of cameras and low-cost radar to perform 3D detection, tracking, and segmentation tasks in Bird's Eye View (BEV) space.
CroCo v2: Improved Cross-view Completion Pre-training for Stereo Matching and Optical Flow
Philippe Weinzaepfel (NAVER LABS Europe), Jerome Revaud (NAVER LABS Europe)
Depth EstimationTransformerOptical FlowImage
🎯 What it does: This paper studies an improved cross-view completion (CroCo v2) pre-training framework aimed at enhancing the performance of dense geometric tasks such as stereo matching and optical flow.
Cross Contrasting Feature Perturbation for Domain Generalization
Chenming Li (Southern University of Science and Technology), Jianguo Zhang (Southern University of Science and Technology)
ClassificationDomain AdaptationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes an online one-stage Cross-Contrast Feature Perturbation framework (CCFP) that simulates domain transfer by applying learnable perturbations to features in the latent space, and introduces a semantic consistency constraint during training to enhance the model's generalization performance on unseen target domains.
Cross Modal Transformer: Towards Fast and Robust 3D Object Detection
Junjie Yan (MEGVII Technology), Xiangyu Zhang (MEGVII Technology)
Object DetectionAutonomous DrivingTransformerImageMultimodalityPoint Cloud
🎯 What it does: An end-to-end cross-modal Transformer (CMT) is proposed, which directly uses image and point cloud tokens to output 3D boxes without explicit perspective transformation, and is robust to LiDAR failures.
Cross-Domain Product Representation Learning for Rich-Content E-Commerce
Xuehan Bai (Kuaishou Technology), Han Li (Kuaishou Technology)
RetrievalRepresentation LearningTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes a large-scale cross-domain e-commerce dataset ROPE, which covers three domains: product pages, short videos, and live broadcasts. It also designs a unified cross-domain product representation framework COPE, using multi-modal contrastive learning and classification loss to learn cross-domain shared features. The framework is then evaluated on cross-domain retrieval and one-shot few-sample classification tasks.
Cross-modal Latent Space Alignment for Image to Avatar Translation
Manuel Ladron de Guevara (Carnegie Mellon University), Daichi Ito (Adobe Research)
Image TranslationGenerationAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: A cross-modal alignment framework is proposed, which generates high-quality, interpretable vectorized avatars (parametric avatars) from a single portrait photo, achieving identity preservation and style transfer.
Cross-Modal Learning with 3D Deformable Attention for Action Recognition
Sangwon Kim (Keimyung University), Byoung Chul Ko (Keimyung University)
RecognitionTransformerVideoMultimodality
🎯 What it does: This paper proposes a Transformer based on 3D deformable attention for fusing video and skeleton modalities, dynamically capturing spatiotemporal features and achieving cross-modal learning, ultimately enabling efficient action recognition.
Cross-Modal Orthogonal High-Rank Augmentation for RGB-Event Transformer-Trackers
Zhiyu Zhu (City University of Hong Kong), Dapeng Oliver Wu (City University of Hong Kong)
Object TrackingTransformerMultimodality
🎯 What it does: This paper proposes two pluggable training enhancement strategies—cross-modal masking modeling and orthogonal high-rank regularization—to improve the cross-modal interaction capability of the pre-trained ViT in the RGB-Event cross-modal tracking task.
Cross-modal Scalable Hyperbolic Hierarchical Clustering
Teng Long (University of Amsterdam), Nanne van Noord (University of Amsterdam)
ClassificationRecognitionRepresentation LearningContrastive LearningVideoMultimodalityAudio
🎯 What it does: This paper proposes a scalable hyperplane hierarchical clustering method (sHHC) that learns continuous hierarchical structures in hyperplane space and applies it to cross-modal self-supervised learning to generate high-quality hierarchical pseudo-labels.
Cross-Modal Translation and Alignment for Survival Analysis
Fengtao Zhou (Hong Kong University of Science and Technology), Hao Chen (Hong Kong University of Science and Technology)
ClassificationData-Centric LearningTransformerMultimodalityBiomedical Data
🎯 What it does: A cross-modal translation and alignment framework (CMTA) is proposed, which jointly predicts the survival time of cancer patients using pathological images and genomic data.
Cross-Ray Neural Radiance Fields for Novel-View Synthesis from Unconstrained Image Collections
Yifan Yang (South China University of Technology), Mingkui Tan (South China University of Technology)
SegmentationGenerationData SynthesisTransformerNeural Radiance FieldImage
🎯 What it does: This study investigates how to recover neural radiance fields from unconstrained image collections (different lighting, shooting times, and transient objects) to achieve novel view synthesis, proposing the Cross-Ray Neural Radiance Field (CR-NeRF).
Cross-view Semantic Alignment for Livestreaming Product Recognition
Wenjie Yang (Kuaishou Technology), Han Li (Kuaishou Technology)
RecognitionRetrievalTransformerContrastive LearningImageVideoTextMultimodality
🎯 What it does: This paper proposes the large-scale multimodal live streaming product recognition and retrieval dataset LPR4M, and designs the RICE framework to address the product recognition and retrieval tasks in live shopping.
Cross-view Topology Based Consistent and Complementary Information for Deep Multi-view Clustering
Zhibin Dong (National University of Defense Technology), En Zhu (National University of Defense Technology)
Auto EncoderImage
🎯 What it does: In multi-view unsupervised clustering, the CTCC framework is proposed, which utilizes cross-view topological graphs to simultaneously learn consistent and complementary information.
CROSSFIRE: Camera Relocalization On Self-Supervised Features from an Implicit Representation
Arthur Moreau (Huawei France), Arnaud de La Fortelle (Mines Paris PSL University)
Pose EstimationRepresentation LearningConvolutional Neural NetworkNeural Radiance FieldImage
🎯 What it does: This paper proposes CROSSFIRE, a self-supervised learning framework that embeds local features into NeRF to achieve real-time camera relocalization using a single RGB image.
CrossLoc3D: Aerial-Ground Cross-Source 3D Place Recognition
Tianrui Guan (University of Maryland), Dinesh Manocha (University of Maryland)
RecognitionRetrievalConvolutional Neural NetworkContrastive LearningPoint CloudBenchmark
🎯 What it does: CROSSLOC3D is proposed, an end-to-end network for cross-source (air-ground) 3D location recognition, utilizing multi-scale voxelization, feature selection, iterative refinement, and NetVLAD to construct global descriptors.
CrossMatch: Source-Free Domain Adaptive Semantic Segmentation via Cross-Modal Consistency Training
Yifang Yin (Institute for Infocomm Research), Roger Zimmermann (National University of Singapore)
SegmentationDomain AdaptationKnowledge DistillationImageMultimodality
🎯 What it does: A source-agnostic domain adaptive semantic segmentation framework is proposed, utilizing a multimodal auxiliary network and cross-modal consistency training to denoise pseudo-labels.
CSDA: Learning Category-Scale Joint Feature for Domain Adaptive Object Detection
Changlong Gao (Xi'an Jiaotong University), Xueming Qian (Xi'an Jiaotong University)
Object DetectionDomain AdaptationGraph Neural NetworkTransformerImage
🎯 What it does: A domain adaptation object detection framework called CSDA is proposed, which achieves category feature alignment and enhancement at different scales.
CTP:Towards Vision-Language Continual Pretraining via Compatible Momentum Contrast and Topology Preservation
Hongguang Zhu (Beijing Jiaotong University), Yao Zhao (Beijing Jiaotong University)
RetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes a new method for Visual-Language Continuous Pre-training (VLCP) called CTP, which addresses the catastrophic forgetting problem of traditional Continual Learning (CIL) under unlabeled and fixed-dimensional embeddings.
CTVIS: Consistent Training for Online Video Instance Segmentation
Kaining Ying (Zhejiang University), Chunhua Shen (Zhejiang University)
Object DetectionSegmentationTransformerContrastive LearningVideo
🎯 What it does: A consistent training method (CTVIS) is proposed, utilizing long video training, a memory bank, momentum average embeddings, and noise injection to align the training process of online video instance segmentation with the inference process, thereby enhancing the distinguishability of instance embeddings and addressing challenges such as occlusion and re-identification.
Cumulative Spatial Knowledge Distillation for Vision Transformers
Borui Zhao (MEGVII Technology), Jiajun Liang (MEGVII Technology)
Knowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: A new knowledge distillation method is proposed - Cumulative Spatial Knowledge Distillation (CSKD), which directly uses the dense spatial predictions of CNN as supervision for ViT, avoiding the difficulties of aligning intermediate features;
CuNeRF: Cube-Based Neural Radiance Field for Zero-Shot Medical Image Arbitrary-Scale Super Resolution
Zixuan Chen (Sun Yat-sen University), Xiaohua Xie (Sun Yat-sen University)
RestorationSuper ResolutionNeural Radiance FieldImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: We propose CuNeRF—a zero-shot medical image arbitrary scale super-resolution (MIASSR) and free-view slice synthesis framework that can generate continuous volume representations solely from low-resolution volumes, enabling the generation of high-quality slices at any magnification and from any viewing angle.
Curvature-Aware Training for Coordinate Networks
Hemanth Saratchandran (Australian Institute of Machine Learning, University of Adelaide), Simon Lucey (Australian Institute of Machine Learning, University of Adelaide)
OptimizationComputational EfficiencyImageVideoAudio
🎯 What it does: This study investigates how to use second-order optimization methods (L-BFGS) to accelerate the training of Coordinate Networks (Coordinate-MLP).
CVRecon: Rethinking 3D Geometric Feature Learning For Neural Reconstruction
Ziyue Feng (Clemson University), Bing Li (Clemson University)
Representation LearningConvolutional Neural NetworkPoint Cloud
🎯 What it does: The CVRecon framework is proposed, utilizing a new Ray-Contextual Compensated Cost Volume (RCCV) to achieve end-to-end 3D neural reconstruction;
CVSformer: Cross-View Synthesis Transformer for Semantic Scene Completion
Haotian Dong (Tianjin University), Di Lin (Tianjin University)
SegmentationTransformerImage
🎯 What it does: A cross-view synthesis Transformer (CVSformer) framework is proposed, which generates multi-view features from a single RGB-D image by rotating 3D convolutional kernels and utilizes a cross-view Transformer to fuse these features for semantic scene completion.
Cyclic Test-Time Adaptation on Monocular Video for 3D Human Mesh Reconstruction
Hyeongjin Nam (Seoul National University), Kyoung Mu Lee (Seoul National University)
Pose EstimationDomain AdaptationConvolutional Neural NetworkVideoMesh
🎯 What it does: A cyclic testing domain adaptation framework called CycleAdapt is proposed, which utilizes HMRNet and MDNet for iterative adaptation on test videos to achieve more accurate 3D human mesh reconstruction.
Cyclic-Bootstrap Labeling for Weakly Supervised Object Detection
Yufei Yin (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
Object DetectionKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A cyclic bootstrapping labeling framework CBL is proposed for weakly supervised object detection, improving pseudo-label generation and enhancing detection performance.
D-IF: Uncertainty-aware Human Digitization via Implicit Distribution Field
Xueting Yang (Psyche AI Inc.), Zhaoxin Fan (Hong Kong University of Science and Technology)
SegmentationGenerationPose EstimationMesh
🎯 What it does: By introducing point-level Gaussian distribution uncertainty to predict human mesh, the implicit field reconstruction method is improved.
D3G: Exploring Gaussian Prior for Temporal Sentence Grounding with Glance Annotation
Hanjun Li (Youtu Lab, Tencent), Xing Sun (Youtu Lab, Tencent)
RetrievalOptimizationTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: A framework D3G is proposed for video sentence localization using glance annotation, aiming to reduce annotation costs while maintaining performance close to fully supervised methods.
DALL-Eval: Probing the Reasoning Skills and Social Biases of Text-to-Image Generation Models
Jaemin Cho (University of North Carolina Chapel Hill), Mohit Bansal (University of North Carolina Chapel Hill)
Object DetectionGenerationTransformerVision Language ModelImage
🎯 What it does: This paper systematically evaluates the visual reasoning capabilities and social biases of text-to-image generation models by designing a new assessment framework.
Dancing in the Dark: A Benchmark towards General Low-light Video Enhancement
Huiyuan Fu (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)
RestorationVideoBenchmark
🎯 What it does: A high-quality dynamic low-light video dataset, DID, is proposed, and an adaptive illumination iterative enhancement network, LAN, is designed based on the Retinex theory for enhancing low-light videos.
DandelionNet: Domain Composition with Instance Adaptive Classification for Domain Generalization
Lanqing Hu (Institute of Computing Technology), Xilin Chen (Institute of Computing Technology)
ClassificationDomain AdaptationConvolutional Neural NetworkImageBenchmark
🎯 What it does: This paper proposes DandelionNet, which utilizes instance-adaptive classifiers (slightly deviating from general classifiers) to achieve a 'combination' of multiple source domains rather than strict alignment, in order to enhance domain generalization capabilities.
Dark Side Augmentation: Generating Diverse Night Examples for Metric Learning
Albert Mohwald (Czech Technical University in Prague), Ondřej Chum (Czech Technical University in Prague)
Image TranslationGenerationRetrievalGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: A method for unpaired nighttime image data augmentation is proposed, which converts daytime images to nighttime images by training a lightweight GAN (HED N GAN) and jointly trains an edge detector to ensure that the edges of the generated nighttime images are consistent with the original images, used for metric learning in image retrieval.
DarSwin: Distortion Aware Radial Swin Transformer
Akshaya Athwale (Université Laval), Jean-François Lalonde (Université Laval)
ClassificationTransformerImage
🎯 What it does: A Transformer structure named DarSwin is proposed, specifically designed for images with known radial distortion (such as wide-angle lenses). It can adaptively process distorted images and perform classification while maintaining polar coordinate patch partitioning, distortion-aware sampling, and angular relative position encoding.
DARTH: Holistic Test-time Adaptation for Multiple Object Tracking
Mattia Segu (ETH Zurich), Fisher Yu (Max Planck Institute for Informatics)
Object TrackingDomain AdaptationConvolutional Neural NetworkContrastive LearningVideo
🎯 What it does: This paper proposes DARTH, a full-process testing-time adaptive framework for multi-object tracking (MOT);
Data Augmented Flatness-aware Gradient Projection for Continual Learning
Enneng Yang (Northeastern University), Xingwei Wang (Northeastern University)
ClassificationOptimizationImage
🎯 What it does: A data augmentation-based Flattened Gradient Projection method (DFGP) is proposed, which optimizes the flatness of the loss surface jointly at the data layer and weight layer to reduce catastrophic forgetting and enhance performance in learning new tasks.
Data-Free Class-Incremental Hand Gesture Recognition
Shubhra Aich (Carnegie Mellon University), Fernando De la Torre (Carnegie Mellon University)
ClassificationRecognitionContrastive LearningPoint Cloud
🎯 What it does: A data-free incremental learning framework is proposed, utilizing Boundary-aware Prototype Model Inversion (BOAT-MI) to achieve zero-shot incremental learning in 3D skeleton gesture recognition.
Data-free Knowledge Distillation for Fine-grained Visual Categorization
Renrong Shao (East China Normal University), Jun Wang (East China Normal University)
ClassificationKnowledge DistillationGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: This paper proposes a data-free knowledge distillation framework for fine-grained visual classification (FGVC) called DFKD-FGVC, which efficiently transfers fine-grained knowledge by combining three key technologies: spatial attention generator, mixed high-order attention distillation (MHAD), and semantic feature contrastive learning (SFCL).
DataDAM: Efficient Dataset Distillation with Attention Matching
Ahmad Sajedi (University of Toronto), Konstantinos N. Plataniotis (University of Toronto)
Data SynthesisKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: Achieve efficient data distillation by learning synthesized images through attention matching.
Dataset Quantization
Daquan Zhou (Bytedance Inc.), Jiashi Feng (Bytedance Inc.)
ClassificationObject DetectionSegmentationCompressionImageText
🎯 What it does: Proposes the Dataset Quantization (DQ) framework, which compresses large-scale datasets into smaller subsets, achieving no significant performance loss across various network architectures while also considering storage and computational efficiency.
DCPB: Deformable Convolution Based on the Poincare Ball for Top-view Fisheye Cameras
Xuan Wei (Southeast University), Xiaobo Lu (Southeast University)
SegmentationConvolutional Neural NetworkGraph Neural NetworkImage
🎯 What it does: For the semantic segmentation task of top fish-eye cameras, this paper proposes a DCPB method for learning deformable convolution kernels in Poincaré spherical space, and integrates it into segmentation networks such as UNet.
DDColor: Towards Photo-Realistic Image Colorization via Dual Decoders
Xiaoyang Kang (DAMO Academy, Alibaba Group), Xuansong Xie (DAMO Academy, Alibaba Group)
Image TranslationRestorationConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: A dual-decoder network DDColor is proposed for unsupervised image colorization.
DDFM: Denoising Diffusion Model for Multi-Modality Image Fusion
Zixiang Zhao (Xi'an Jiaotong University), Luc Van Gool (ETH Zurich)
GenerationData SynthesisDiffusion modelImageMultimodalityBiomedical DataMagnetic Resonance ImagingComputed TomographyPositron Emission Tomography
🎯 What it does: A multi-modal image fusion method DDFM based on denoising diffusion models is proposed, combining unsupervised diffusion generation with EM-style likelihood correction.
DDG-Net: Discriminability-Driven Graph Network for Weakly-supervised Temporal Action Localization
Xiaojun Tang (Beijing University of Posts and Telecommunications), Zongyuan Yang (Beijing University of Posts and Telecommunications)
RecognitionObject DetectionGraph Neural NetworkOptical FlowVideo
🎯 What it does: This paper proposes DDG-Net, a graph network specifically designed for weakly supervised temporal action localization. It enhances the representation of video segments by classifying them into three categories: pseudo actions, pseudo backgrounds, and ambiguous segments, and employs different graph connection methods to improve the distinguishability of segment representations, further enhancing localization performance.
DDIT: Semantic Scene Completion via Deformable Deep Implicit Templates
Haoang Li (Technical University of Munich), Daniel Cremers (Technical University of Munich)
SegmentationGenerationOptimizationPoint CloudMesh
🎯 What it does: In a three-dimensional scene, after performing semantic segmentation on each instance, instance-level semantic scene completion is achieved using a deformable deep implicit template (a variant of DeepSDF) combined with predicted latent codes.
DDP: Diffusion Model for Dense Visual Prediction
Yuanfeng Ji (University of Hong Kong), Ping Luo (Huawei Noah's Ark Lab)
SegmentationDepth EstimationAutonomous DrivingTransformerDiffusion modelImage
🎯 What it does: A dense visual prediction framework DDP based on conditional diffusion models is proposed, which unifies the handling of tasks such as semantic segmentation, BEV segmentation, and depth estimation.
DDS2M: Self-Supervised Denoising Diffusion Spatio-Spectral Model for Hyperspectral Image Restoration
Yuchun Miao (Wuhan University), Dacheng Tao (University of Sydney)
RestorationDiffusion modelImage
🎯 What it does: A self-supervised diffusion-based spatiotemporal spectral model DDS2M is proposed for recovering clear images from degraded hyperspectral images.
Dec-Adapter: Exploring Efficient Decoder-Side Adapter for Bridging Screen Content and Natural Image Compression
Sheng Shen (Tianjin University), Jingyu Yang (Tianjin University)
CompressionDomain AdaptationOptimizationConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: This study investigates a lightweight decoder adapter that addresses the significant performance drop when transferring natural image compression models to screen content images.
DECO: Dense Estimation of 3D Human-Scene Contact In The Wild
Shashank Tripathi (Max Planck Institute for Intelligent Systems), Michael J. Black (Max Planck Institute for Intelligent Systems)
SegmentationPose EstimationTransformerImage
🎯 What it does: This work proposes a model called DECO that can predict dense 3D contact between the human body surface and objects or scenes from a single outdoor image, and has collected a large-scale dense contact annotation dataset called DAMON.
Decomposition-Based Variational Network for Multi-Contrast MRI Super-Resolution and Reconstruction
Pengcheng Lei (East China Normal University), Tieyong Zeng (Chinese University of Hong Kong)
RestorationSuper ResolutionConvolutional Neural NetworkImageMultimodalityMagnetic Resonance Imaging
🎯 What it does: A multi-contrast MRI super-resolution and reconstruction network based on a variational model has been constructed, which can decompose the reference image into common and unique components, and only pass the common information to the target image;
Decouple Before Interact: Multi-Modal Prompt Learning for Continual Visual Question Answering
Zi Qian (Tsinghua University), Wenwu Zhu (Tsinghua University)
ClassificationRecognitionTransformerPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A comprehensive framework for Continuous Visual Question Answering (CL-VQA) is proposed, along with the TRIPLET method, which achieves experience-free continual learning through decoupled multimodal prompt learning on a pretrained vision-language model.
Decoupled DETR: Spatially Disentangling Localization and Classification for Improved End-to-End Object Detection
Manyuan Zhang (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
Object DetectionTransformerImage
🎯 What it does: This paper proposes a spatially decoupled DETR (SD-DETR), which splits the cross-attention blocks in the decoder and shares self-attention to generate task-aware queries for classification and localization tasks, thereby achieving spatial decoupling of features and predictions, and introduces alignment loss to bridge the inconsistency between high-confidence classification and precise localization.