CVPR 2023 Papers with AI Summaries
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2353 papers
→ CVPR 2023 papers with code (830)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 2353 CVPR 2023 papers by keyword, author or institution —
free trial on arXivSub.
"Seeing" Electric Network Frequency From Events
Lexuan Xu (Wuhan University), Ning Qiao (SynSense Tech Co Ltd)
Video
🎯 What it does: A method for estimating power grid frequency (ENF) based on event cameras, called E-ENF, is proposed, and an event-video hybrid ENF dataset is constructed.
(ML)$^2$P-Encoder: On Exploration of Channel-Class Correlation for Multi-Label Zero-Shot Learning
Ziming Liu (Hong Kong Polytechnic University), Fushuo Huo (Hong Kong Polytechnic University)
ClassificationRecognitionConvolutional Neural NetworkImage
🎯 What it does: A channel-category related multi-label zero-shot learning framework C-MLZSL is proposed, and a lightweight (ML)P-Encoder 2 module is designed to achieve channel semantic extraction and fusion.
1% VS 100%: Parameter-Efficient Low Rank Adapter for Dense Predictions
Dongshuo Yin (Chinese Academy of Sciences), Xian Sun (Chinese Academy of Sciences)
Object DetectionSegmentationTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes LoRand, a low-rank adapter designed for efficiently fine-tuning large-scale visual models in dense prediction tasks, compressing trainable parameters to 1-3% while maintaining performance close to full fine-tuning.
1000 FPS HDR Video With a Spike-RGB Hybrid Camera
Yakun Chang (Peking University), Boxin Shi (Peking University)
RestorationData SynthesisRecurrent Neural NetworkSpiking Neural NetworkOptical FlowVideo
🎯 What it does: A hybrid camera system combining a neuromorphic synaptic camera and an alternating exposure RGB camera has been constructed, and an end-to-end framework has been proposed to generate HDR video at 1000 frames per second;
2PCNet: Two-Phase Consistency Training for Day-to-Night Unsupervised Domain Adaptive Object Detection
Mikhail Kennerley (National University of Singapore), Robby T. Tan (National University of Singapore)
Object DetectionDomain AdaptationAutonomous DrivingConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a two-stage consistency learning framework, 2PCNet, for unsupervised domain adaptation in object detection during day and night.
3D Cinemagraphy From a Single Image
Xingyi Li (Huazhong University of Science and Technology), Guosheng Lin (Nanyang Technological University)
GenerationData SynthesisDepth EstimationGenerative Adversarial NetworkOptical FlowImageVideoPoint Cloud
🎯 What it does: This paper proposes a complete framework for generating 3D Cinemagraphs (visual dynamic images with camera motion) from a single static image, integrating image animation and single-image novel view synthesis.
3D Concept Learning and Reasoning From Multi-View Images
Yining Hong (University of California Los Angeles), Chuang Gan (University of Massachusetts Amherst)
Object DetectionSegmentationRepresentation LearningConvolutional Neural NetworkRecurrent Neural NetworkVision Language ModelContrastive LearningImage
🎯 What it does: The paper proposes a new task called 3DMV-VQA for 3D visual question answering from multi-view images and constructs a large-scale dataset consisting of approximately 5k scenes, 600k images, and 50k question-answer pairs.
3D GAN Inversion With Facial Symmetry Prior
Fei Yin (Shenzhen International Graduate School Tsinghua University), Yujiu Yang (Tencent AI Lab)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a 3D GAN inversion method based on facial symmetry priors, enabling the recovery of high-quality, view-consistent 3D facial geometry and texture from a single image.
3D Highlighter: Localizing Regions on 3D Shapes via Text Descriptions
Dale Decatur, Rana Hanocka
Object DetectionSegmentationVision Language ModelMesh
🎯 What it does: This study proposes 3D Highlighter, a technique for semantic region localization of 3D meshes based on text descriptions;
3D Human Keypoints Estimation From Point Clouds in the Wild Without Human Labels
Zhenzhen Weng (Stanford University), Dragomir Anguelov (Waymo)
SegmentationPose EstimationDomain AdaptationAutonomous DrivingTransformerOptical FlowPoint Cloud
🎯 What it does: This paper proposes a completely unsupervised 3D human keypoint detection method (GC-KPL) that can learn human joint positions in outdoor LiDAR point clouds without the need for manually annotated 3D keypoints.
3D Human Mesh Estimation From Virtual Markers
Xiaoxuan Ma (Peking University), Yizhou Wang (Peking University)
Pose EstimationConvolutional Neural NetworkMesh
🎯 What it does: Proposes and learns 'Virtual Marker' as an intermediate representation for 3D human mesh estimation, capable of predicting 3D marker points from a single image and recovering the complete mesh through interpolation.
3D Human Pose Estimation via Intuitive Physics
Shashank Tripathi (Max Planck Institute for Intelligent Systems), Dimitrios Tzionas (University of Amsterdam)
Pose EstimationImage
🎯 What it does: This paper proposes a 3D human pose estimation method based on Intuitive Physics (IP) called IPMAN. By incorporating the center of mass (CoM), pressure heatmaps, and center of pressure (CoP) into the loss function, the method achieves a more stable and physically plausible 3D SMPL model reconstruction from a single frame image.
3D Human Pose Estimation With Spatio-Temporal Criss-Cross Attention
Zhenhua Tang (Hefei University of Technology), Ting Yao (University of Science and Technology of China)
Pose EstimationTransformerImageVideo
🎯 What it does: This paper proposes a Transformer-based framework for 2D to 3D human pose estimation—STCFormer, which efficiently models spatiotemporal correlations using a Spatiotemporal Cross Attention (STC) module.
3D Line Mapping Revisited
Shaohui Liu (ETH Zurich), Viktor Larsson (Lund University)
Object DetectionOptimizationSimultaneous Localization and MappingImage
🎯 What it does: A complete 3D line segment mapping pipeline, LIMAP, is proposed for reconstructing high-quality 3D line segments from multi-view images and generating line-point and line VP association maps.
3D Neural Field Generation Using Triplane Diffusion
J. Ryan Shue (Milton Academy), Gordon Wetzstein (Stanford University)
GenerationData SynthesisDiffusion modelPoint CloudMesh
🎯 What it does: Utilizing a 2D diffusion model to generate a 3D neural field, which first converts ShapeNet meshes into occupancy fields and decomposes them into triplane features, then trains the diffusion model on these 2D feature maps, generating triplanes during inference and decoding them into 3D shapes through a shared MLP;
3D Registration With Maximal Cliques
Xiyu Zhang (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
Pose EstimationOptimizationPoint Cloud
🎯 What it does: A 3D point cloud registration method based on maximal clique (MAC) is proposed, which generates pose hypotheses by searching for maximal cliques in the compatibility graph.
3D Semantic Segmentation in the Wild: Learning Generalized Models for Adverse-Condition Point Clouds
Aoran Xiao (Nanyang Technological University), Eric P. Xing
SegmentationDomain AdaptationContrastive LearningPoint Cloud
🎯 What it does: A large-scale extreme weather LiDAR point cloud dataset called SemanticSTF is proposed, along with the design of the PointDR framework based on geometric style randomization and contrastive learning for learning semantic segmentation models under all weather conditions.
3D Shape Reconstruction of Semi-Transparent Worms
Thomas P. Ilett (University of Leeds), David C. Hogg (University of Leeds)
OptimizationVideo
🎯 What it does: This paper proposes an end-to-end method based on a differentiable renderer and curve parameterization to achieve 3D shape reconstruction of translucent nematodes (Caenorhabditis elegans) in gel under multi-view from three cameras.
3D Spatial Multimodal Knowledge Accumulation for Scene Graph Prediction in Point Cloud
Mingtao Feng (Xidian University), Ajmal Mian (University of Western Australia)
Object DetectionSegmentationGenerationGraph Neural NetworkTransformerMultimodalityPoint CloudGraph
🎯 What it does: Generating semantic scene graphs from 3D point clouds, utilizing spatial hierarchical knowledge to assist in relationship prediction.
3D Video Loops From Asynchronous Input
Li Ma (Hong Kong University of Science and Technology), Pedro V. Sander (Hong Kong University of Science and Technology)
GenerationOptimizationComputational EfficiencyNeural Radiance FieldVideo
🎯 What it does: Using completely asynchronous multi-view short videos, a multi-tile video representation (Multi-tile Video, MTV) is constructed that can loop infinitely.
3D Video Object Detection With Learnable Object-Centric Global Optimization
Jiawei He (Chinese Academy of Sciences), Zhaoxiang Zhang (Chinese Academy of Sciences)
Object DetectionPose EstimationAutonomous DrivingOptimizationVideoPoint Cloud
🎯 What it does: A two-stage 3D video object detection framework called BA-Det is proposed, which jointly optimizes object detection and pose estimation using learnable long-term visual correspondence and optimizable object-centric bundle adjustment.
3D-Aware Conditional Image Synthesis
Kangle Deng (Carnegie Mellon University), Jun-Yan Zhu (Carnegie Mellon University)
GenerationData SynthesisNeural Radiance FieldGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes Pix2Pix3D, a 3D generative model that maps 2D label maps (such as segmentation maps or edge maps) to renderable 3D neural fields, capable of synthesizing high-quality images from arbitrary viewpoints and generating corresponding pixel-aligned label maps.
3D-Aware Face Swapping
Yixuan Li (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
Image TranslationGenerationGenerative Adversarial NetworkImage
🎯 What it does: A novel facial swapping method called 3dSwap based on 3D GAN is proposed, which can achieve high-fidelity and multi-view consistent facial overlap between single-view source and target images.
3D-Aware Facial Landmark Detection via Multi-View Consistent Training on Synthetic Data
Libing Zeng (Texas A&M University), Nima Khademi Kalantari
Data SynthesisPose EstimationNeural Radiance FieldImage
🎯 What it does: A 3D consistency training method for facial keypoint localization is proposed, utilizing a synthetic multi-view dataset and self-projection consistency loss, along with a pluggable 3D perception module to enhance the localization accuracy of existing networks.
3D-Aware Multi-Class Image-to-Image Translation With NeRFs
Senmao Li (Nankai University), Jian Yang (Nankai University)
Image TranslationGenerationNeural Radiance FieldGenerative Adversarial NetworkImage
🎯 What it does: Proposes a 3D-aware multi-class image-to-image (I2I) translation method based on NeRF.
3D-Aware Object Goal Navigation via Simultaneous Exploration and Identification
Jiazhao Zhang (Peking University), He Wang (Peking University)
Robotic IntelligenceReinforcement LearningPoint Cloud
🎯 What it does: A 3D perception-based target object navigation framework is proposed, which utilizes online point cloud construction and fusion to generate a dense 3D semantic scene, and simultaneously executes a corner-point guided exploration strategy and a category-adaptive recognition strategy within the same framework, thereby completing the object target navigation task in unknown environments.
3D-POP - An Automated Annotation Approach to Facilitate Markerless 2D-3D Tracking of Freely Moving Birds With Marker-Based Motion Capture
Hemal Naik (Max Planck Institute of Animal Behavior), Máté Nagy (Hungarian Academy of Sciences)
Object TrackingPose EstimationSupervised Fine-TuningVideo
🎯 What it does: This paper proposes a semi-automatic labeling method based on a motion capture (Mo-Cap) system, constructing the first 3D pose and identity labeling dataset for birds (pigeons) called 3D-POP, which includes approximately 300,000 frames and 2 million keypoint instances, covering multiple angles, multiple individuals, and various behaviors.
3DAvatarGAN: Bridging Domains for Personalized Editable Avatars
Rameen Abdal (King Abdullah University of Science and Technology), Sergey Tulyakov (Snap Inc.)
GenerationDomain AdaptationGenerative Adversarial NetworkImage
🎯 What it does: This paper studies a domain adaptation framework based on 3D GAN, which can train a 3D avatar generator on artistic style data without camera annotations, achieving personalized and editable 3D avatars from a single image.
3Mformer: Multi-Order Multi-Mode Transformer for Skeletal Action Recognition
Lei Wang (Australian National University), Piotr Koniusz (Data61 CSIRO)
RecognitionTransformerVideo
🎯 What it does: The 3Mformer model is proposed, which utilizes high-order Transformer (HoT) and hypergraph structures to encode skeletal actions, and integrates them through multi-order multi-modal Transformer to enhance skeletal action recognition performance.
A Bag-of-Prototypes Representation for Dataset-Level Applications
Weijie Tu (Australian National University), Liang Zheng (Australian National University)
Representation LearningData-Centric LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A Bag-of-Prototypes (BoP) dataset representation method is proposed, which uses clustering to obtain a prototype codebook and statistics the projection histograms of image features on these prototypes to form dataset vectors.
A Characteristic Function-Based Method for Bottom-Up Human Pose Estimation
Haoxuan Qu (Singapore University of Technology and Design), Jun Liu (Singapore University of Technology and Design)
Pose EstimationConvolutional Neural NetworkImage
🎯 What it does: A training method for optimizing the underlying human pose estimation model is proposed by minimizing the distance of the feature function between the predicted heatmap and the ground truth heatmap.
A Data-Based Perspective on Transfer Learning
Saachi Jain (Massachusetts Institute of Technology), Aleksander Mądry (Massachusetts Institute of Technology)
Domain AdaptationData-Centric LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: A data-driven framework is proposed to measure the impact of various classes (or samples) in the source dataset on the performance of downstream tasks in transfer learning, and to improve performance by removing negative classes/samples.
A Dynamic Multi-Scale Voxel Flow Network for Video Prediction
Xiaotao Hu (Nankai University), Shuchang Zhou (Megvii Technology)
Autonomous DrivingOptical FlowVideo
🎯 What it does: A Dynamic Multi-Scale Voxel Flow Network (DMVFN) is proposed to perform video prediction using only RGB frames.
A General Regret Bound of Preconditioned Gradient Method for DNN Training
Hongwei Yong (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
Object DetectionOptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: A general regret upper bound for constrained full matrix preconditioning gradients is proposed, and based on this theory, the AdaBK optimizer is designed (which can be embedded into SGDM and AdamW, resulting in SGDM BK and AdamW BK).
A Generalized Framework for Video Instance Segmentation
Miran Heo (Yonsei University), Seon Joo Kim (Yonsei University)
Object DetectionSegmentationTransformerVideo
🎯 What it does: A general video instance segmentation framework GenVIS is proposed, capable of processing long videos in online and semi-online modes.
A Hierarchical Representation Network for Accurate and Detailed Face Reconstruction From In-the-Wild Images
Biwen Lei (Alibaba Group), Xuansong Xie (Alibaba Group)
RestorationGenerationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A Hierarchical Representation Network (HRN) is proposed, capable of achieving high-precision and detail-rich 3D face reconstruction from a single image, which can be further extended to multi-view reconstruction.
A Large-Scale Homography Benchmark
Daniel Barath (ETH Zurich), Jiri Matas (Czech Technical University in Prague)
Pose EstimationOptimizationSupervised Fine-TuningImageBenchmark
🎯 What it does: A large planar dataset Pi3D (approximately 1000 3D planes) and corresponding 226,260 ground truth homography sub-images (HEB) have been constructed, and they are used to systematically evaluate various robust estimators and deep pre-filtering methods.
A Large-Scale Robustness Analysis of Video Action Recognition Models
Madeline Chantry Schiappa (University of Central Florida), Yogesh S. Rawat
RecognitionConvolutional Neural NetworkTransformerVideoBenchmark
🎯 What it does: This paper conducts a large-scale evaluation of the robustness of existing video action recognition models under real distribution shifts.
A Light Touch Approach to Teaching Transformers Multi-View Geometry
Yash Bhalgat (Visual Geometry Group University of Oxford), Andrew Zisserman (Visual Geometry Group University of Oxford)
RetrievalTransformerContrastive LearningImage
🎯 What it does: This paper proposes an object retrieval re-ranking method that implicitly guides the Transformer to learn multi-view geometric knowledge through epipolar constraints (Epipolar Loss). During the training phase, camera pose information is used to construct epipolar lines, guiding attention to focus on features along the corresponding epipolar lines, while inference does not require any camera information.
A Light Weight Model for Active Speaker Detection
Junhua Liao (Sichuan University), Liangyin Chen (Sichuan University)
RecognitionComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkMultimodalityBenchmarkAudio
🎯 What it does: A lightweight real-time active speaker detection framework is proposed, using a single candidate input.
A Loopback Network for Explainable Microvascular Invasion Classification
Shengxuming Zhang (Zhejiang University), Mingli Song (Zhejiang University)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: Proposes LoopNet, a dual-branch recurrent network that utilizes only image-level labels to classify microvascular invasion (MVI) and locate cancer cells;
A Meta-Learning Approach to Predicting Performance and Data Requirements
Achin Jain (AWS AI Labs), Stefano Soatto (University of Southern California)
ClassificationObject DetectionMeta LearningTabular
🎯 What it does: A piece of research proposes a piecewise power law model (PPL) based on meta-learning to predict model performance and the required amount of data from few-sample data.
A New Benchmark: On the Utility of Synthetic Data With Blender for Bare Supervised Learning and Downstream Domain Adaptation
Hui Tang (South China University of Technology), Kui Jia (South China University of Technology)
Data SynthesisDomain AdaptationConvolutional Neural NetworkTransformerImageBenchmark
🎯 What it does: A large-scale synthetic image dataset is generated using Blender and domain randomization techniques, systematically evaluating the utility of synthetic data in supervised learning and subsequent domain adaptation, and proposing a more challenging S2RDA evaluation benchmark.
A New Comprehensive Benchmark for Semi-Supervised Video Anomaly Detection and Anticipation
Congqi Cao (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
Anomaly DetectionAuto EncoderVideoBenchmark
🎯 What it does: A novel large-scale multi-scene semi-supervised video anomaly detection and prediction dataset, NWPU Campus, is proposed, and a model based on forward-backward frame prediction and scene conditional variational autoencoder is developed to achieve both anomaly detection and early warning.
A New Dataset Based on Images Taken by Blind People for Testing the Robustness of Image Classification Models Trained for ImageNet Categories
Reza Akbarian Bafghi (University of Colorado), Danna Gurari (University of Colorado)
ClassificationImage
🎯 What it does: A new dataset named VizWiz-Classification for images taken by blind individuals has been constructed, containing 8,900 images, 200 ImageNet categories, and annotated with multi-labels and quality defects; this dataset is used to evaluate the robustness of 100 ImageNet classification models.
A New Path: Scaling Vision-and-Language Navigation With Synthetic Instructions and Imitation Learning
Aishwarya Kamath (New York University), Zarana Parekh (Google Research)
TransformerReinforcement LearningVision Language ModelText
🎯 What it does: A large-scale dataset containing 4.2 million synthetic navigation instructions and trajectories was constructed, and state-of-the-art performance was achieved on the RxR task using a Transformer agent based solely on imitation learning.
A Practical Stereo Depth System for Smart Glasses
Jialiang Wang (Meta Platforms Inc), Matt Uyttendaele (Meta Platforms Inc)
Depth EstimationConvolutional Neural NetworkImage
🎯 What it does: The paper designs and implements an end-to-end stereo depth perception system for smart glasses, including preprocessing, online stereo calibration, depth estimation, and a monocular backup scheme, and utilizes the depth results for 3D computational photography effects.
A Practical Upper Bound for the Worst-Case Attribution Deviations
Fan Wang (Nanyang Technological University), Adams Wai-Kin Kong (Nanyang Technological University)
OptimizationExplainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a theoretical limit on the attribution invariance of upper bound evaluation models for feature attribution methods in deep networks when subjected to perturbations.
A Probabilistic Attention Model With Occlusion-Aware Texture Regression for 3D Hand Reconstruction From a Single RGB Image
Zheheng Jiang (Lancaster University), Bryan M. Williams (Lancaster University)
RestorationPose EstimationConvolutional Neural NetworkTransformerImage
🎯 What it does: A probability-based attention model is proposed, combining the MANO prior with AMVUR to achieve 3D hand reconstruction and texture recovery from a single RGB image, supporting both supervised and weakly supervised training.
A Probabilistic Framework for Lifelong Test-Time Adaptation
Dhanajit Brahma (Indian Institute of Technology Kanpur), Piyush Rai (Indian Institute of Technology Kanpur)
Domain AdaptationKnowledge DistillationImage
🎯 What it does: A probabilistic framework for Lifelong Test-Time Adaptation (TTA) called PETAL is proposed, which utilizes the posterior distribution of the source domain as a prior, constructs a student-teacher cross-entropy loss, and incorporates regularization.
A Rotation-Translation-Decoupled Solution for Robust and Efficient Visual-Inertial Initialization
Yijia He (Chinese Academy of Sciences), Hongdong Li (Australian National University)
Pose EstimationOptimizationComputational EfficiencySimultaneous Localization and MappingVideo
🎯 What it does: This paper proposes a visual-inertial initialization method that decouples rotation and translation, achieving fast and high-precision initialization by combining gyroscope bias estimation and linear translation constraints.
A Simple Baseline for Video Restoration With Grouped Spatial-Temporal Shift
Dasong Li (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
RestorationConvolutional Neural NetworkVideo
🎯 What it does: A lightweight video restoration framework based on grouped spatial-temporal shifts is designed to achieve video deblurring and denoising without using optical flow, deformable convolution, or self-attention.
A Simple Framework for Text-Supervised Semantic Segmentation
Muyang Yi (Shanghai Jiao Tong University), Hongtao Lu (Shanghai Jiao Tong University)
SegmentationTransformerPrompt EngineeringContrastive LearningImageText
🎯 What it does: Proposes the SimSeg framework, which utilizes a pre-trained CLIP model combined with text supervision to achieve zero-shot semantic segmentation.
A Soma Segmentation Benchmark in Full Adult Fly Brain
Xiaoyu Liu (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
Object DetectionSegmentationConvolutional Neural NetworkImageBiomedical DataBenchmark
🎯 What it does: A two-stage deep learning method was designed for the instance segmentation of cell bodies (soma) from high-resolution EM data of the complete adult fruit fly brain, and the first cell body annotation dataset (EMADS) was constructed.
A Strong Baseline for Generalized Few-Shot Semantic Segmentation
Sina Hajimiri (École de Technologie Supérieure), Jose Dolz
SegmentationKnowledge DistillationImage
🎯 What it does: A general few-shot semantic segmentation framework called DIaM is proposed. This framework uses standard supervised learning during the training phase and employs optimization methods that maximize mutual information and knowledge distillation for incremental few-shot adaptation of any pre-trained segmentation network during the inference phase.
A Unified HDR Imaging Method With Pixel and Patch Level
Qingsen Yan (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
RestorationConvolutional Neural NetworkTransformerImage
🎯 What it does: A novel HDR de-ghosting network called HyHDRNet is proposed, which integrates patch aggregation, pixel-level ghost attention, and transformer fusion to efficiently remove ghosts caused by motion and overexposure in dynamic scenes while restoring details.
A Unified Knowledge Distillation Framework for Deep Directed Graphical Models
Yizhuo Chen (William and Mary), Huajie Shao (William and Mary)
CompressionKnowledge DistillationGraph Neural NetworkAuto EncoderImage
🎯 What it does: A unified knowledge distillation framework is proposed for deep directed graph models, utilizing reparameterization to convert all latent variables into deterministic auxiliary variables, achieving model compression and continuous learning without training data.
A Unified Pyramid Recurrent Network for Video Frame Interpolation
Xin Jin (Samsung Electronics), Cheul-hee Hahm (Samsung Electronics)
RestorationData SynthesisRecurrent Neural NetworkOptical FlowVideo
🎯 What it does: A lightweight unified pyramid recursive network, UPR-Net, is proposed for video frame interpolation, utilizing forward warping and iterative refinement to achieve high-quality intermediate frame synthesis.
A Unified Spatial-Angular Structured Light for Single-View Acquisition of Shape and Reflectance
Xianmin Xu (Zhejiang University), Hongzhi Wu (Zhejiang University)
Depth EstimationOptimizationAuto EncoderPoint CloudMesh
🎯 What it does: A unified spatial-angular structured light system is designed, utilizing an LED array and an LCD mask to simultaneously capture the 3D shape and SVBRDF of objects in a single view with high quality.
A Whac-a-Mole Dilemma: Shortcuts Come in Multiples Where Mitigating One Amplifies Others
Zhiheng Li (University of Rochester), Mark Ibrahim (Meta AI)
ClassificationAutonomous DrivingImageBenchmark
🎯 What it does: To address the phenomenon of multiple shortcuts, this paper proposes two new benchmark datasets (UrbanCars and ImageNet-W), systematically evaluates the limitations of existing shortcut mitigation methods, and introduces the Last Layer Ensemble (LLE) method to jointly suppress various shortcuts.
A-Cap: Anticipation Captioning With Commonsense Knowledge
Duc Minh Vo, Hideki Nakayama
GenerationGraph Neural NetworkTransformerContrastive LearningImageVideoTextMultimodalityTime Series
🎯 What it does: This paper proposes a temporal sparse image autoregressive model A-CAP based on graph neural networks and Transformers for predicting future descriptive captions (anticipation captioning) given only a very small number of time series images.
A-La-Carte Prompt Tuning (APT): Combining Distinct Data via Composable Prompting
Benjamin Bowman (AWS AI Labs), Stefano Soatto (AWS AI Labs)
ClassificationData-Centric LearningTransformerPrompt EngineeringImage
🎯 What it does: A pluggable prompt tuning method called A‑la‑carte Prompt Tuning (APT) is proposed, achieving composable sub-data models to meet the needs of continual learning, data forgetting, and user customization;
A2J-Transformer: Anchor-to-Joint Transformer Network for 3D Interacting Hand Pose Estimation From a Single RGB Image
Changlong Jiang (Huazhong University of Science and Technology), Joey Tianyi Zhou (Agency for Science Technology and Research, A*STAR)
Pose EstimationTransformerImage
🎯 What it does: Proposes the A2J-Transformer model, which utilizes 3D anchors and Transformer to achieve interactive hand pose estimation from a single RGB image.
ABCD: Arbitrary Bitwise Coefficient for De-Quantization
Woo Kyoung Han (Daegu Gyeongbuk Institute of Science and Technology), Kyong Hwan Jin (Daegu Gyeongbuk Institute of Science and Technology)
RestorationImageVideo
🎯 What it does: An arbitrary bit-depth expansion (ABCD) model based on implicit neural representation is proposed, capable of recovering high bit-depth images from any quantized low bit-depth images, and supports multiple bit depths simultaneously during training.
ABLE-NeRF: Attention-Based Rendering With Learnable Embeddings for Neural Radiance Field
Zhe Jun Tang (Nanyang Technological University), Haiyu Zhao (SenseTime Research)
TransformerNeural Radiance FieldImage
🎯 What it does: A Transformer-based ABLE-NeRF is designed, utilizing self-attention to simulate volumetric rendering and incorporating learnable embeddings to memorize scene lighting, thereby enhancing perspective-dependent rendering quality.
Abstract Visual Reasoning: An Algebraic Approach for Solving Raven's Progressive Matrices
Jingyi Xu (Singapore University of Technology and Design), Kai Fong Ernest Chong (Singapore University of Technology and Design)
Object DetectionGenerationImage
🎯 What it does: This paper proposes an 'algebraic machine reasoning' framework based on algebraic ideals, abstracting Raven’s Progressive Matrices (RPM) problems as ideal operations in polynomial rings, utilizing algebraic subroutines such as Gröbner bases and primary decomposition for answer selection and generation.
Accelerated Coordinate Encoding: Learning to Relocalize in Minutes Using RGB and Poses
Eric Brachmann (Niantic), Victor Adrian Prisacariu (Niantic)
Pose EstimationComputational EfficiencySimultaneous Localization and MappingImage
🎯 What it does: Rapidly train the visual relocalization system ACE, completing new scene mapping in just 5 minutes while maintaining accuracy comparable to DSAC*.
Accelerating Dataset Distillation via Model Augmentation
Lei Zhang (Zhejiang University), Dongkuan Xu (North Carolina State University)
ClassificationComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: An efficient dataset distillation method based on model augmentation is proposed, which accelerates the traditional gradient matching distillation process by using early-trained models and parameter perturbations.
Accelerating Vision-Language Pretraining With Free Language Modeling
Teng Wang (Southern University of Science and Technology), Ping Luo (Southern University of Science and Technology)
RetrievalComputational EfficiencyRepresentation LearningTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Proposed the Free Language Modeling (FLM) objective and the encode-corrupt-predict framework to accelerate visual-language pre-training.
AccelIR: Task-Aware Image Compression for Accelerating Neural Restoration
Juncheol Ye (Korea Advanced Institute of Science and Technology), Dongsu Han (Korea Advanced Institute of Science and Technology)
RestorationSuper ResolutionCompressionConvolutional Neural NetworkImage
🎯 What it does: An adaptive compression framework called AccelIR is proposed for image restoration tasks, which can enhance the quality of the final restoration network and reduce computational load while keeping the image compression size unchanged through adaptive adjustment of block-level quantization parameters (QP).
Accidental Light Probes
Hong-Xing Yu (Stanford University), Deqing Sun (Google Research)
Pose EstimationOptimizationImage
🎯 What it does: Using everyday smooth objects (such as soda cans, thermoses, etc.) as 'accidental light probes', we reconstruct their geometric shapes and spatially varying BRDFs offline, and use differentiable rendering to inversely solve their poses and environmental lighting from a single image, achieving high-fidelity estimation of scene lighting and object insertion/re-lighting.
Achieving a Better Stability-Plasticity Trade-Off via Auxiliary Networks in Continual Learning
Sanghwan Kim (ETH Zurich), Thomas Hofmann (ETH Zurich)
OptimizationKnowledge DistillationImage
🎯 What it does: A framework is proposed that utilizes an auxiliary network to enhance the stability-plasticity balance in continual learning.
ACL-SPC: Adaptive Closed-Loop System for Self-Supervised Point Cloud Completion
Sangmin Hong (Seoul National University), Kyoung Mu Lee (Seoul National University)
GenerationData SynthesisPoint Cloud
🎯 What it does: A fully self-supervised point cloud completion framework, ACL-SPC, is proposed, which uses a single incomplete point cloud as input to generate a complete 3D point cloud.
ACR: Attention Collaboration-Based Regressor for Arbitrary Two-Hand Reconstruction
Zhengdi Yu (Tencent AI Lab), Jue Wang (Tencent AI Lab)
Pose EstimationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an Attention Collaboration-based Regressor (ACR) that can achieve 3D reconstruction of any two hands from a single RGB image, taking into account complex situations such as hand interactions, truncation, and occlusion.
ACSeg: Adaptive Conceptualization for Unsupervised Semantic Segmentation
Kehan Li (Peking University), Jie Chen (Peking University)
SegmentationTransformerContrastive LearningImage
🎯 What it does: This paper proposes an adaptive conceptualization framework called ACSeg based on self-supervised ViT features for unsupervised semantic segmentation.
Actionlet-Dependent Contrastive Learning for Unsupervised Skeleton-Based Action Recognition
Lilang Lin (Peking University), Jiaying Liu (Peking University)
RecognitionRepresentation LearningGraph Neural NetworkContrastive LearningVideoMultimodality
🎯 What it does: For unsupervised skeleton action recognition, the paper proposes a contrastive learning framework based on actionlets, utilizing motion-adaptive data transformations and semantic-aware feature pooling to enhance representation quality.
Activating More Pixels in Image Super-Resolution Transformer
Xiangyu Chen (University of Macau), Chao Dong (Shenzhen Institute of Advanced Technology)
RestorationSuper ResolutionTransformerImage
🎯 What it does: This paper proposes a Hybrid Attention Transformer (HAT) framework for single image super-resolution tasks.
Active Exploration of Multimodal Complementarity for Few-Shot Action Recognition
Yuyang Wanyan (Chinese Academy of Sciences), Changsheng Xu (University of Science and Technology of China)
ClassificationRecognitionKnowledge DistillationMeta LearningOptical FlowVideoMultimodality
🎯 What it does: A multi-modal few-shot action recognition framework based on active learning (AMFAR) is proposed, which improves the robustness of few-shot action classification by identifying reliable modalities and adaptively fusing them in each task.
Active Finetuning: Exploiting Annotation Budget in the Pretraining-Finetuning Paradigm
Yichen Xie (University of California), Wei Zhan (University of California)
ClassificationSegmentationOptimizationTransformerSupervised Fine-TuningImage
🎯 What it does: This paper defines the Active Finetuning task, which aims to select a small subset of samples from a large pool of unlabeled data for supervised finetuning, and proposes the ActiveFT method to achieve this goal.
ActMAD: Activation Matching To Align Distributions for Test-Time-Training
Muhammad Jehanzeb Mirza (Institute for Computer Graphics and Vision), Horst Bischof (Institute for Computer Graphics and Vision)
Object DetectionDomain AdaptationAutonomous DrivingImage
🎯 What it does: This paper proposes an online test-time training method called ActMAD based on location-aware activation matching to adapt to distribution shifts in the absence of source data.
AdaMAE: Adaptive Masking for Efficient Spatiotemporal Learning With Masked Autoencoders
Wele Gedara Chaminda Bandara (Johns Hopkins University), Vishal M. Patel (Zippin)
RecognitionComputational EfficiencyTransformerReinforcement LearningAuto EncoderVideo
🎯 What it does: An end-to-end trainable adaptive masking strategy called AdaMAE has been developed for video masked autoencoders, achieving efficient spatiotemporal learning.
AdamsFormer for Spatial Action Localization in the Future
Hyung-gun Chi (Purdue University), Chiho Choi (Samsung Semiconductor)
Convolutional Neural NetworkTransformerVideoOrdinary Differential Equation
🎯 What it does: This paper proposes the Future Frame Action Localization (SALF) task and constructs the AdamsFormer model based on the multi-step Adams method of Neural ODE, which can predict and locate future actions from observed frames.
Adapting Shortcut With Normalizing Flow: An Efficient Tuning Framework for Visual Recognition
Yaoming Wang, Qi Tian
ClassificationObject DetectionSegmentationRepresentation LearningContrastive LearningImage
🎯 What it does: This paper proposes a representation learning framework based on Conditional Mutual Information (CMI), which aims to obtain more discriminative features by maximizing the CMI between the representation and the original data.
Adaptive Annealing for Robust Geometric Estimation
Chitturi Sidhartha (Indian Institute of Science), Venu Madhav Govindu (Indian Institute of Science)
Autonomous DrivingOptimizationPoint Cloud
🎯 What it does: A GNC method based on Hessian positive definiteness adaptive annealing is proposed for robust geometric estimation, particularly for 3D point cloud registration.
Adaptive Assignment for Geometry Aware Local Feature Matching
Dihe Huang (Tsinghua University), Chengjie Wang (Tencent YouTu Lab)
Pose EstimationDepth EstimationTransformerImage
🎯 What it does: This paper proposes AdaMatcher, an end-to-end detector-free feature matching framework that provides adaptive assignment and scale alignment.
Adaptive Channel Sparsity for Federated Learning Under System Heterogeneity
Dongping Liao (University of Macau), Cheng-Zhong Xu
Federated LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a federated learning method called Flado based on adaptive channel sparsification, which significantly improves model convergence speed and final accuracy under system heterogeneity.
Adaptive Data-Free Quantization
Biao Qian (Hefei University of Technology), Meng Wang (Hefei University of Technology)
ClassificationOptimizationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: Proposes Adaptive Data-Free Quantization (AdaDFQ), which calibrates the quantization network by generating adaptively adjustable synthetic samples without the need for original training data.
Adaptive Global Decay Process for Event Cameras
Urbano Miguel Nunes (Sorbonne University), Sio-Hoi Ieng (Sorbonne University)
ClassificationObject DetectionConvolutional Neural NetworkOptical FlowImageVideo
🎯 What it does: An adaptive global decay process is proposed for event stream processing of event cameras, and its effectiveness is validated across various tasks.
Adaptive Graph Convolutional Subspace Clustering
Lai Wei (Shanghai Maritime University), Jin Liu (Shanghai Maritime University)
Graph Neural NetworkImage
🎯 What it does: Adaptive Graph Convolutional Subspace Clustering (AGCSC) is achieved by using graph convolution technology combined with feature extraction and self-expression matrix constraints.
Adaptive Human Matting for Dynamic Videos
Chung-Ching Lin (Microsoft), Zicheng Liu (Microsoft)
SegmentationTransformerVideo
🎯 What it does: The AdaM framework is proposed to achieve alpha blending of human foregrounds in dynamic videos without the need for trimaps or pre-collected backgrounds.
Adaptive Patch Deformation for Textureless-Resilient Multi-View Stereo
Yuesong Wang (Huazhong University of Science and Technology), Yawei Luo (Zhejiang University)
Depth EstimationOptimizationImageBenchmark
🎯 What it does: An adaptive patch deformation is introduced based on traditional PatchMatch, dynamically expanding the receptive field for unreliable pixels in texture-missing areas, and using anchor pixels to ensure that the matching cost approaches global optimality, thereby achieving texture-insensitive and low-memory multi-view stereo reconstruction.
Adaptive Plasticity Improvement for Continual Learning
Yan-Shuo Liang (Nanjing University), Wu-Jun Li (Nanjing University)
ClassificationImage
🎯 What it does: An Adaptive Plasticity Improvement (API) framework is proposed for continual learning, addressing catastrophic forgetting and model plasticity.
Adaptive Sparse Convolutional Networks With Global Context Enhancement for Faster Object Detection on Drone Images
Bowei Du (Beihang University), Di Huang (Beihang University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: A method for accelerating detection heads based on sparse convolution, called CEASC, is proposed for object detection in drone images.
Adaptive Sparse Pairwise Loss for Object Re-Identification
Xiao Zhou (Tsinghua University), Lin Ma (Meituan Inc.)
RecognitionRetrievalConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: This paper proposes a Sparse Pairwise (SP) loss and its adaptive version AdaSP for the task of person/vehicle re-identification, where only one pair of positive samples and one pair of negative samples are sampled for each category during training.
Adaptive Spot-Guided Transformer for Consistent Local Feature Matching
Jiahuan Yu (University of Science and Technology of China), Feng Wu (China Academy of Space Technology)
TransformerImage
🎯 What it does: Proposes the Adaptive Spot-guided Transformer (ASTR) for consistent local feature matching.
Adaptive Zone-Aware Hierarchical Planner for Vision-Language Navigation
Chen Gao (Beihang University), Si Liu (Beihang University)
Robotic IntelligenceTransformerReinforcement LearningVision-Language-Action ModelMultimodality
🎯 What it does: This paper proposes the Adaptive Zone-aware Hierarchical Planner (AZHP), which achieves hierarchical planning in visual language navigation through adaptive zone partitioning (SZP) and goal-oriented zone selection (GZS), combined with a state switching module (SSM) to enable asynchronous switching between high-level and low-level actions.
AdaptiveMix: Improving GAN Training via Feature Space Shrinkage
Haozhe Liu (King Abdullah University of Science and Technology), Yefeng Zheng (Tencent)
GenerationAnomaly DetectionConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A pluggable AdaptiveMix module is proposed, which enhances the training stability of GANs and the quality of generated images by pulling mixed samples closer in the feature space of the discriminator, thereby shrinking the distribution area of training samples.
Adjustment and Alignment for Unbiased Open Set Domain Adaptation
Wuyang Li (City University of Hong Kong), Yixuan Yuan (The Chinese University of Hong Kong)
Domain AdaptationImage
🎯 What it does: To address the semantic-level bias problem in open set domain adaptation, the ANNA framework is proposed, which discovers potential new category regions in the source domain through Front-Door Adjustment (FDA) to achieve unbiased learning. It then separates and aligns the base classes and new classes through Decoupled Causal Alignment (DCA), ultimately achieving unbiased open set domain adaptation.
Advancing Visual Grounding With Scene Knowledge: Benchmark and Method
Zhihong Chen (Chinese University of Hong Kong), Guanbin Li (Sun Yat-sen University)
RecognitionObject DetectionTransformerVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This work proposes the Scene Knowledge Driven Visual Grounding (SK-VG) task and constructs a corresponding dataset; it also designs two knowledge embedding-based single-stage methods, KeViLI, and a language structure-based two-stage method, LeViLM, to explore the role of scene knowledge in visual-language alignment.
Adversarial Counterfactual Visual Explanations
Guillaume Jeanneret (University of Caen Normandie), Frédéric Jurie (University of Caen Normandie)
Autonomous DrivingExplainability and InterpretabilityAdversarial AttackDiffusion modelImage
🎯 What it does: The ACE method is proposed, which generates semantic and executable adversarial counterfactual images through adversarial attacks to explain the decisions of the target model.