CVPR 2023 Papers — Page 19
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2353 papers
Revisiting Prototypical Network for Cross Domain Few-Shot Learning
Fei Zhou (Northwestern Polytechnical University), Yanning Zhang (University of Wollongong)
Domain AdaptationKnowledge DistillationImage
🎯 What it does: This paper proposes a local-global knowledge distillation framework LDP-net to enhance the generalization performance of prototype networks in cross-domain few-shot learning.
Revisiting Residual Networks for Adversarial Robustness
Shihua Huang (Michigan State University), Vishnu Naresh Boddeti (Michigan State University)
ClassificationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: The system evaluates the impact of different architectural designs of residual networks on adversarial robustness and proposes a new residual block called RobustResBlock, a wide-narrow composite scaling rule called RobustScaling, and RobustResNets based on both.
Revisiting Reverse Distillation for Anomaly Detection
Tran Dinh Tien (VinBrain JSC), Steven Q. H. Truong (VinUniversity)
Anomaly DetectionKnowledge DistillationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Improved the Reverse Distillation (RD) framework in unsupervised anomaly detection by incorporating a pseudo-anomaly mechanism, multi-scale projection layers, and multi-task learning to enhance feature compactness and suppress anomaly information.
Revisiting Rolling Shutter Bundle Adjustment: Toward Accurate and Fast Solution
Bangyan Liao (Hunan University), Yizhen Lao (Hunan University)
Pose EstimationOptimizationComputational EfficiencySimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes a Normalized and Covariance Weighted Rolling Shutter Bundle Adjustment (NW-RSBA) that can achieve high-precision and efficient pose and 3D point optimization for rolling shutter camera data without motion constraints or frame rate limitations.
Revisiting Rotation Averaging: Uncertainties and Robust Losses
Ganlin Zhang (ETH Zurich), Daniel Barath (ETH Zurich)
Pose EstimationOptimizationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes to directly use the covariance uncertainty of two-view geometry for rotation averaging in global SfM to improve camera pose estimation.
Revisiting Self-Similarity: Structural Embedding for Image Retrieval
Seongwon Lee (Yonsei University), Euntai Kim (Yonsei University)
RetrievalConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A new global image retrieval embedding network called SENet is proposed, which can simultaneously encode visual features and internal structural information in an end-to-end learning manner.
Revisiting Temporal Modeling for CLIP-Based Image-to-Video Knowledge Transferring
Ruyang Liu (Peking University), Thomas H. Li (Peking University)
RecognitionRetrievalTransformerContrastive LearningVideoText
🎯 What it does: This paper proposes a branch structure called Spatial-Temporal Auxiliary Network (STAN) to transfer image-text pre-trained models like CLIP to video tasks, addressing the issue of simultaneously utilizing high-level semantic knowledge and low-level visual patterns.
Revisiting the P3P Problem
Yaqing Ding (Lund University), Kalle Åström (Lund University)
Pose EstimationImage
🎯 What it does: This paper proposes a P3P solver based on the intersection of two conjugate cones. By analyzing the geometric significance of the roots of a cubic equation, specialized solving strategies are developed for different root configurations, achieving fast and numerically stable camera pose estimation.
Revisiting the Stack-Based Inverse Tone Mapping
Ning Zhang (Peking University), Ronggang Wang (Peking University)
RestorationConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: A method for reconstructing HDR images from a single LDR image using only two generated exposure images is proposed.
Revisiting Weak-to-Strong Consistency in Semi-Supervised Semantic Segmentation
Lihe Yang (Southeast University), Yinghuan Shi (Southeast University)
SegmentationContrastive LearningImageBiomedical Data
🎯 What it does: In the semi-supervised semantic segmentation task, the weak-strong consistency framework of FixMatch is migrated and further improved.
RGB No More: Minimally-Decoded JPEG Vision Transformers
Jeongsoo Park (University of Michigan), Justin Johnson (University of Michigan)
CompressionComputational EfficiencyTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes training the Vision Transformer (ViT) directly on the DCT series of JPEG encoding, without fully decoding to RGB, thereby significantly reducing data loading costs and improving training/inference speed.
RGBD2: Generative Scene Synthesis via Incremental View Inpainting Using RGBD Diffusion Models
Jiabao Lei (South China University of Technology), Kui Jia (South China University of Technology)
RestorationGenerationData SynthesisDiffusion modelImagePoint CloudMesh
🎯 What it does: Based on sparse RGBD views, new RGBD images are recursively generated along the camera trajectory, and a RGBD diffusion model is used to jointly fill in the missing areas of the rendered images and depth. Finally, all completed views are projected back onto a 3D mesh and fused to obtain the complete scene geometry and color.
RIATIG: Reliable and Imperceptible Adversarial Text-to-Image Generation With Natural Prompts
Han Liu (Washington University in St. Louis), Ning Zhang (Washington University in St. Louis)
GenerationAdversarial AttackLarge Language ModelGenerative Adversarial NetworkImageText
🎯 What it does: A reliable and covert targeted adversarial attack method for text-image generation models, named RIATIG, is proposed, which can generate natural prompts that are semantically aligned with the target image but have almost no relevance to the text content.
RIAV-MVS: Recurrent-Indexing an Asymmetric Volume for Multi-View Stereo
Changjiang Cai (OPPO US Research Center), Yi Xu (OPPO US Research Center)
Depth EstimationOptimizationRecurrent Neural NetworkTransformerPoint Cloud
🎯 What it does: A learning-optimized multi-view stereo matching framework RIAV-MVS is proposed, which iteratively optimizes depth through recursive indexing of asymmetric cost volumes.
RIDCP: Revitalizing Real Image Dehazing via High-Quality Codebook Priors
Rui-Qi Wu (Nankai University), Chongyi Li (Nanyang Technological University)
RestorationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: A real image dehazing network based on high-quality codebook priors (RIDCP) is proposed, and a controllable HQPs matching and more realistic fog synthesis pipeline is designed.
RIFormer: Keep Your Vision Backbone Effective but Removing Token Mixer
Jiahao Wang (Shanghai AI Laboratory), Dahua Lin (Shanghai AI Laboratory)
ClassificationComputational EfficiencyKnowledge DistillationTransformerImage
🎯 What it does: A visual backbone network RIFormer is proposed, which does not require a token mixer and achieves efficient inference through structural re-parameterization and knowledge distillation.
Rigidity-Aware Detection for 6D Object Pose Estimation
Yang Hai (Xidian University), Yinlin Hu (EPFL)
Object DetectionPose EstimationSupervised Fine-TuningImageBenchmark
🎯 What it does: A detection method based on rigid object visibility guidance is proposed to improve 2D bounding box detection in scenes with severe occlusion for 6D object pose estimation.
RILS: Masked Visual Reconstruction in Language Semantic Space
Shusheng Yang (Huazhong University of Science and Technology), Xinggang Wang (Huazhong University of Science and Technology)
RecognitionObject DetectionSegmentationTransformerAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: This paper studies a visual pre-training framework that unifies Masked Image Modeling (MIM) with language supervision (CLIP), proposing masked visual reconstruction in the language semantic space (RILS);
RMLVQA: A Margin Loss Approach for Visual Question Answering With Language Biases
Abhipsa Basu (Indian Institute of Science), R. Venkatesh Babu (Indian Institute of Science)
RecognitionOptimizationContrastive LearningImageMultimodality
🎯 What it does: A robust adaptive angular margin loss (RMLVQA) is proposed to suppress language bias in visual question answering and achieve better generalization without using data augmentation.
Robot Structure Prior Guided Temporal Attention for Camera-to-Robot Pose Estimation From Image Sequence
Yang Tian (Peking University), Hao Dong (Peking University)
Pose EstimationRobotic IntelligenceImageVideo
🎯 What it does: Using robot structural priors and a temporal attention mechanism, online pose estimation between the camera and the robot base is achieved based on continuous RGB frames.
Robust 3D Shape Classification via Non-Local Graph Attention Network
Shengwei Qin (Zhejiang Sci-Tech University), Ligang Liu (University of Science and Technology of China)
ClassificationGraph Neural NetworkPoint Cloud
🎯 What it does: A non-local graph attention network (NLGAT) is designed to capture global relationships and structures through multi-scale Gram matrices, achieving robust 3D shape classification for rotated, sparse, and noisy point clouds.
Robust and Scalable Gaussian Process Regression and Its Applications
Yifan Lu (Wuhan University), Junjun Jiang (Hunan University)
Anomaly DetectionOptimizationGaussian SplattingTabularBiomedical Data
🎯 What it does: A robust and scalable Gaussian Process Regression (GPR) model based on variational learning is proposed to handle large-scale data containing outliers.
Robust Dynamic Radiance Fields
Yu-Lun Liu (National Taiwan University), Jia-Bin Huang (Meta)
Data SynthesisPose EstimationNeural Radiance FieldOptical FlowVideo
🎯 What it does: This paper proposes a neural radiance field framework, RoDynRF, that can directly reconstruct dynamic scenes from monocular videos without pre-defined camera poses, jointly optimizing camera intrinsic and extrinsic parameters, as well as static and dynamic radiance fields.
Robust Generalization Against Photon-Limited Corruptions via Worst-Case Sharpness Minimization
Zhuo Huang (University of Sydney), Tongliang Liu (Nanjing University of Science and Technology)
OptimizationImage
🎯 What it does: The SharpDRO method is proposed to enhance robust generalization against photon-limited noise by minimizing sharpness on the worst-case distribution.
Robust Mean Teacher for Continual and Gradual Test-Time Adaptation
Mario Döbler (University of Stuttgart), Bin Yang (University of Stuttgart)
Domain AdaptationContrastive LearningImage
🎯 What it does: A robust mean teacher (RMT) framework is proposed for test-time adaptation (TTA) during continuous and progressive testing, continuously improving model performance post-deployment through self-supervised learning.
Robust Model-Based Face Reconstruction Through Weakly-Supervised Outlier Segmentation
Chunlu Li (Southeast University), Adam Kortylewski (Max Planck Institute for Informatics)
RestorationSegmentationGenerationConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: This paper proposes a weakly supervised 3D face reconstruction framework called FOCUS, which achieves high-quality reconstruction in occluded environments by jointly training a face autoencoder and an out-of-bounds segmentation network.
Robust Multiview Point Cloud Registration With Reliable Pose Graph Initialization and History Reweighting
Haiping Wang (Wuhan University), Bisheng Yang (Wuhan University)
Pose EstimationOptimizationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes a multi-view point cloud registration method that combines learning-based sparse graph construction with historical reweighting iterative least squares (IRLS).
Robust Outlier Rejection for 3D Registration With Variational Bayes
Haobo Jiang (Nanjing University of Science and Technology), Mathieu Salzmann (EPFL)
Pose EstimationAnomaly DetectionAutonomous DrivingOptimizationRecurrent Neural NetworkPoint Cloud
🎯 What it does: A non-local network based on variational Bayesian inference is proposed for filtering outliers in 3D point cloud registration, along with a voting-based inlier search method.
Robust Single Image Reflection Removal Against Adversarial Attacks
Zhenbo Song (Nanjing University of Science and Technology), Jianfeng Lu (Nanjing University of Science and Technology)
RestorationAdversarial AttackTransformerImage
🎯 What it does: This paper focuses on the task of single image reflection removal (SIRR), studying its robustness against adversarial attacks and proposing a new robust network.
Robust Test-Time Adaptation in Dynamic Scenarios
Longhui Yuan (Beijing Institute of Technology), Shuang Li (Beijing Institute of Technology)
Domain AdaptationImageBenchmark
🎯 What it does: Proposes a Practical Test-Time Adaptation (PTTA) method for distribution shifts and time-related sampling called RoTTA.
Robust Unsupervised StyleGAN Image Restoration
Yohan Poirier-Ginter (Inria), Jean-François Lalonde (University Laval)
RestorationGenerative Adversarial NetworkImage
🎯 What it does: A StyleGAN inverse image restoration framework is proposed that does not require separate parameter tuning for each level of denoising or task; robust recovery for various degradation tasks such as upsampling, denoising, artifact removal, and inpainting is achieved through a three-stage gradual expansion of the latent space and the use of conservative normalized gradient descent optimization.
RobustNeRF: Ignoring Distractors With Robust Losses
Sara Sabour (Google Research), Andrea Tagliasacchi (Google Research)
Data SynthesisOptimizationNeural Radiance FieldImage
🎯 What it does: This study investigates the artifact problem caused by non-persistent disturbances such as moving objects and shadows during the training process of NeRF, and proposes RobustNeRF, which automatically ignores these disturbances through robust loss and trimmed least squares methods.
RODIN: A Generative Model for Sculpting 3D Digital Avatars Using Diffusion
Tengfei Wang (Hong Kong University of Science and Technology), Baining Guo (Microsoft Research)
GenerationData SynthesisDiffusion modelNeural Radiance FieldImage
🎯 What it does: This work proposes a 3D diffusion model RODIN, capable of automatically generating high-fidelity digital avatars (NeRF) and supporting generation from a single portrait or text prompts, as well as text-based semantic editing.
Role of Transients in Two-Bounce Non-Line-of-Sight Imaging
Siddharth Somasundaram (Massachusetts Institute of Technology), Ramesh Raskar (Massachusetts Institute of Technology)
RestorationDepth EstimationSimultaneous Localization and MappingImage
🎯 What it does: This paper proposes the use of two-hop transients for non-line-of-sight (NLOS) imaging, combined with multiplexed illumination to significantly reduce the required number of measurements and improve reconstruction quality.
RONO: Robust Discriminative Learning With Noisy Labels for 2D-3D Cross-Modal Retrieval
Yanglin Feng (Sichuan University), Peng Hu (Institute for Infocomm Research A*STAR)
RetrievalContrastive LearningMultimodality
🎯 What it does: A robust learning framework RONO for 2D-3D cross-modal retrieval is proposed, which can effectively learn cross-modal discrimination and invariant representation under the condition of noisy labels.
Rotation-Invariant Transformer for Point Cloud Matching
Hao Yu (TU Munich), Slobodan Ilic (TU Munich)
TransformerPoint Cloud
🎯 What it does: This paper proposes a Transformer model called RoITr that maintains rotation invariance under arbitrary poses for point cloud matching.
Run, Don't Walk: Chasing Higher FLOPS for Faster Neural Networks
Jierun Chen (Hong Kong University of Science and Technology), S.-H. Gary Chan (Hong Kong University of Science and Technology)
ClassificationObject DetectionSegmentationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A new operator called Partial Convolution (PConv) is proposed, and based on this operator, the FasterNet network is constructed, aiming to enhance the network's FLOPS (i.e., actual computation speed) while reducing FLOPs, thus achieving faster neural network inference.
RUST: Latent Neural Scene Representations From Unposed Imagery
Mehdi S. M. Sajjadi (Google Research), Klaus Greff (Google Research)
GenerationData SynthesisPose EstimationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a scene representation and novel view synthesis method called RUST, which utilizes implicit pose encoding of the target image to learn a latent representation of 3D scenes and perform viewpoint rendering.
RWSC-Fusion: Region-Wise Style-Controlled Fusion Network for the Prohibited X-Ray Security Image Synthesis
Luwen Duan (Zhejiang Dahua Technology Co), Xi Li (Zhejiang University)
Object DetectionGenerationData SynthesisConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: The Regional Style Control Fusion Network (RWSC-Fusion) automatically overlays prohibited objects onto X-ray security images using an edge attention module, generating realistic synthetic X-ray images with corresponding annotations.
S3C: Semi-Supervised VQA Natural Language Explanation via Self-Critical Learning
Wei Suo (Northwestern Polytechnical University), Qi Wu (University of Adelaide)
Explainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextMultimodality
🎯 What it does: Using prompts and self-criticism reinforcement learning to generate natural language explanations in visual question answering
SadTalker: Learning Realistic 3D Motion Coefficients for Stylized Audio-Driven Single Image Talking Face Animation
Wenxuan Zhang (Xi'an Jiaotong University), Fei Wang (Xi'an Jiaotong University)
GenerationData SynthesisGenerative Adversarial NetworkImageVideoAudio
🎯 What it does: This paper presents the SadTalker system, which can generate realistic and synchronized talking head videos using a single portrait image and audio.
Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models
Patrick Schramowski (German Research Center for Artificial Intelligence), Kristian Kersting (German Research Center for Artificial Intelligence)
GenerationDiffusion modelImage
🎯 What it does: A Safe Latent Diffusion (SLD) method has been designed to actively suppress inappropriate content during the diffusion generation process, and an I2P (Inappropriate Image Prompt) dataset has been proposed to evaluate this method.
Sample-Level Multi-View Graph Clustering
Yuze Tan (Sichuan University), Jiancheng Lv (Sichuan University)
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proposes a multi-view graph clustering method that integrates multi-view information at the sample level and learns topological correlations.
Samples With Low Loss Curvature Improve Data Efficiency
Isha Garg (Purdue University), Kaushik Roy (Purdue University)
OptimizationData-Centric LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: The study investigated the second-order curvature of the loss for training deep networks with respect to input samples, finding that low-curvature samples are highly correlated with the 'cleanliness' of the data, and based on this, proposed the SLo-Curves core set selection and training method.
Sampling Is Matter: Point-Guided 3D Human Mesh Reconstruction
Jeonghwan Kim (Konkuk University), Wonjun Kim (Konkuk University)
GenerationPose EstimationTransformerMesh
🎯 What it does: A 3D human mesh reconstruction method based on a single RGB image is proposed, with the core idea of enhancing the network's ability to capture vertex positions through point-guided feature sampling and progressive attention masking.
SAP-DETR: Bridging the Gap Between Salient Points and Queries-Based Transformer Detector for Fast Model Convergency
Yang Liu (Institute of Computing Technology), Zhiqiang He (Lenovo)
Object DetectionTransformerImage
🎯 What it does: This paper proposes SAP-DETR, which views object detection as a transformation from salient points to bounding boxes, using query-specific salient points as reference points.
SCADE: NeRFs from Space Carving With Ambiguity-Aware Depth Estimates
Mikaela Angelina Uy (Stanford University), Ke Li (Simon Fraser University)
GenerationDepth EstimationOptimizationConvolutional Neural NetworkNeural Radiance FieldImagePoint Cloud
🎯 What it does: Proposes the SCADE method, which optimizes NeRF using multi-modal monocular depth priors through spatial exclusion loss under a sparse and unconstrained perspective, to achieve high-quality 3D reconstruction and novel view synthesis.
Scalable, Detailed and Mask-Free Universal Photometric Stereo
Satoshi Ikehata (National Institute of Informatics)
TransformerImage
🎯 What it does: This paper proposes a general photometric camera method that can directly generate high-resolution surface normal maps under unlit calibration and spatially varying lighting environments.
ScaleDet: A Scalable Multi-Dataset Object Detector
Yanbei Chen (AWS AI Labs), Davide Modolo (AWS AI Labs)
Object DetectionPrompt EngineeringImage
🎯 What it does: A scalable multi-dataset object detector ScaleDet is proposed, which can unify the label space of different datasets during training and can be directly used on any given dataset during inference.
ScaleFL: Resource-Adaptive Federated Learning With Heterogeneous Clients
Fatih Ilhan (Georgia Institute of Technology), Ling Liu (Georgia Institute of Technology)
Federated LearningKnowledge DistillationImageText
🎯 What it does: A resource-adaptive federated learning framework called ScaleFL is proposed, which can dynamically scale the global model in width and depth according to the computational budget of each client and generate multiple exit sub-models, supporting local training and global aggregation under heterogeneous resources.
ScaleKD: Distilling Scale-Aware Knowledge in Small Object Detector
Yichen Zhu (Midea Group), Jian Tang (Midea Group)
Object DetectionKnowledge DistillationImage
🎯 What it does: A scale-aware knowledge distillation method called ScaleKD is proposed to enhance the detection performance of small object detectors without increasing inference costs.
Scaling Language-Image Pre-Training via Masking
Yanghao Li (Meta AI), Kaiming He (Meta AI)
RetrievalComputational EfficiencyTransformerContrastive LearningImageTextMultimodality
🎯 What it does: A method called FLIP is proposed to randomly mask image (optional text) blocks during CLIP training to achieve sparse computation and significantly accelerate training.
Scaling Up GANs for Text-to-Image Synthesis
Minguk Kang (POSTECH), Taesung Park (Adobe Research)
GenerationData SynthesisGenerative Adversarial NetworkImageText
🎯 What it does: GigaGAN is proposed, a GAN architecture scalable to a billion parameters for text-to-image generation, achieving high-resolution image synthesis on large-scale open datasets.
ScanDMM: A Deep Markov Model of Scanpath Prediction for 360deg Images
Xiangjie Sui (Jiangxi University of Finance and Economics), Zhou Wang (University of Waterloo)
GenerationData SynthesisRecurrent Neural NetworkImage
🎯 What it does: A scanning path prediction model for 360° images, ScanDMM, is proposed, which uses a deep Markov model to model the time-dependent dynamics of visual attention and generate realistic gaze trajectories.
ScarceNet: Animal Pose Estimation With Scarce Annotations
Chen Li (National University of Singapore), Gim Hee Lee (National University of Singapore)
Pose EstimationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: Proposes ScarceNet, which performs animal pose estimation using a small number of labeled and unlabeled images through pseudo-labeling, re-labeling with reusable samples, and student-teacher consistency constraints;
SCConv: Spatial and Channel Reconstruction Convolution for Feature Redundancy
Jiafeng Li (East China Normal University), Lianghua He (Tongji University)
Object DetectionSegmentationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A pluggable convolution module SCConv is proposed to simultaneously suppress spatial and channel redundancy in CNNs, improving feature expression efficiency.
Scene-Aware Egocentric 3D Human Pose Estimation
Jian Wang (Max Planck Institute Informatics), Christian Theobalt (Max Planck Institute Informatics)
Pose EstimationDepth EstimationConvolutional Neural NetworkImage
🎯 What it does: Proposes a framework for 3D human pose estimation within the perspective of a single-lens head-mounted fisheye camera using scene constraints;
SceneComposer: Any-Level Semantic Image Synthesis
Yu Zeng (Johns Hopkins University), Vishal M. Patel (Johns Hopkins University)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: A unified conditional image generation framework is proposed, capable of synthesizing high-quality images based on semantic layouts of any precision (from plain text to fine segmentation maps).
SceneTrilogy: On Human Scene-Sketch and Its Complementarity With Photo and Text
Pinaki Nath Chowdhury (University of Surrey), Yi-Zhe Song (University of Surrey)
GenerationRetrievalTransformerContrastive LearningImageTextMultimodality
🎯 What it does: This paper constructs a joint embedding of three modalities: scene sketches, text, and images, allowing for arbitrary combinations for querying or task usage, completing multi-tasks such as retrieval, title generation, and subjective title generation.
SCoDA: Domain Adaptive Shape Completion for Real Scans
Yushuang Wu (Chinese University of Hong Kong Shenzhen), Xiaoguang Han (Chinese University of Hong Kong Shenzhen)
Domain AdaptationPoint CloudMesh
🎯 What it does: This study investigates the transfer of synthetic data knowledge to the task of completing 3D shapes from real scans.
SCOOP: Self-Supervised Correspondence and Optimization-Based Scene Flow
Itai Lang (Tel Aviv University), Michael Rubinstein (Google Research)
Autonomous DrivingOptimizationOptical FlowPoint Cloud
🎯 What it does: The SCOOP method is proposed, which combines self-supervised point cloud correspondence learning and runtime flow refinement to estimate scene flow.
Score Jacobian Chaining: Lifting Pretrained 2D Diffusion Models for 3D Generation
Haochen Wang (TTI-Chicago), Greg Shakhnarovich (TTI-Chicago)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Using a pre-trained 2D diffusion model, we map 2D image gradients to 3D voxel representations through the chain rule and a differentiable renderer, achieving 3D generation without 3D training data.
SCOTCH and SODA: A Transformer Video Shadow Detection Framework
Lihao Liu (University of Cambridge), Angelica I. Aviles-Rivero
Object DetectionSegmentationTransformerContrastive LearningVideo
🎯 What it does: This paper proposes a Transformer-based video shadow detection framework SCOTCH and SODA, which can accurately segment shadow regions in videos.
SCPNet: Semantic Scene Completion on Point Cloud
Zhaoyang Xia (Shanghai AI Laboratory), Yu Qiao (Shanghai AI Laboratory)
SegmentationAutonomous DrivingKnowledge DistillationConvolutional Neural NetworkPoint Cloud
🎯 What it does: A unified semantic scene completion framework, SCPNet, is proposed for sparse and locally missing outdoor LiDAR point clouds, capable of simultaneously performing semantic segmentation and voxel-level semantic completion.
SDC-UDA: Volumetric Unsupervised Domain Adaptation Framework for Slice-Direction Continuous Cross-Modality Medical Image Segmentation
Hyungseob Shin (Yonsei University), Dosik Hwang (Yonsei University)
SegmentationDomain AdaptationConvolutional Neural NetworkGenerative Adversarial NetworkImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: This paper proposes the SD-C-UDA framework for unsupervised domain adaptation of cross-modal medical image volumetric segmentation, achieving high-precision segmentation with continuity in the slice direction.
SDFusion: Multimodal 3D Shape Completion, Reconstruction, and Generation
Yen-Chi Cheng (University of Illinois Urbana-Champaign), Liang-Yan Gui
GenerationData SynthesisDiffusion modelMultimodalityPoint Cloud
🎯 What it does: A SDFusion framework based on diffusion models is proposed, utilizing SDF as a 3D representation and supporting multi-modal conditions (partial shapes, images, text) to generate high-resolution 3D shapes;
SE-ORNet: Self-Ensembling Orientation-Aware Network for Unsupervised Point Cloud Shape Correspondence
Jiacheng Deng (University of Science and Technology of China), Zhe Zhang (China Academy of Space Technology)
Domain AdaptationKnowledge DistillationRepresentation LearningGraph Neural NetworkPoint Cloud
🎯 What it does: A self-ensemble orientation-aware network (SE-ORNet) for unsupervised point cloud shape correspondence is proposed, achieving alignment and robust feature learning of point clouds through a self-ensemble framework and orientation estimation module, thereby obtaining accurate dense correspondences.
Search-Map-Search: A Frame Selection Paradigm for Action Recognition
Mingjun Zhao (University of Alberta), Di Niu (University of Alberta)
RecognitionTransformerVideo
🎯 What it does: A Search-Map-Search (SMS) framework is proposed, which identifies the optimal frame combination through hierarchical search and trains a feature mapping network, enabling the rapid recovery of the optimal frame set during inference, thereby enhancing action recognition performance.
Seasoning Model Soups for Robustness to Adversarial and Natural Distribution Shifts
Francesco Croce (University of Tübingen), Sven Gowal (DeepMind)
Domain AdaptationAdversarial AttackSupervised Fine-TuningImage
🎯 What it does: This paper studies the merging of models that are robust to different ℓp attacks through linear interpolation in the parameter space (model soups), achieving adjustable robustness and enabling rapid adaptation to new distribution shifts with very few samples.
SeaThru-NeRF: Neural Radiance Fields in Scattering Media
Deborah Levy (University of Haifa), Tali Treibitz (University of Haifa)
RestorationGenerationNeural Radiance FieldImage
🎯 What it does: A new perspective rendering NeRF method called SeaThru-NeRF is proposed for scattering media (such as underwater), which can separate objects from scattered light and generate realistic views with or without media.
SECAD-Net: Self-Supervised CAD Reconstruction by Learning Sketch-Extrude Operations
Pu Li (Institute of Automation Chinese Academy of Sciences), Dong-Ming Yan (Institute of Automation Chinese Academy of Sciences)
GenerationConvolutional Neural NetworkPoint CloudMesh
🎯 What it does: This paper proposes a self-supervised SECAD-Net for reconstructing editable CAD models from raw 3D geometry (voxel grids).
Seeing a Rose in Five Thousand Ways
Yunzhi Zhang (Stanford University), Jiajun Wu (Stanford University)
GenerationData SynthesisNeural Radiance FieldGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a model for learning the intrinsic properties of objects (geometry, texture, and material) from a single image containing multiple instances, and utilizes this distribution to generate new instances with different poses and lighting.
Seeing Beyond the Brain: Conditional Diffusion Model With Sparse Masked Modeling for Vision Decoding
Zijiao Chen (National University of Singapore), Juan Helen Zhou (National University of Singapore)
GenerationData SynthesisRepresentation LearningTransformerDiffusion modelBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposes the MinD-Vis two-stage framework: first, self-supervised pre-training is conducted on large-scale unlabeled fMRI using Sparse Mask Brain Modeling (SC-MBM), and then the obtained sparse high-dimensional representation is used as conditional input to generate high-quality visual images corresponding to fMRI using a Dual Conditional Latent Diffusion Model (DC-LDM).
Seeing Through the Glass: Neural 3D Reconstruction of Object Inside a Transparent Container
Jinguang Tong (Australian National University), Hongdong Li (Australian National University)
Object DetectionSegmentationData SynthesisNeural Radiance FieldImage
🎯 What it does: A new problem is proposed for recovering the 3D geometry of objects within transparent containers, and the ReNeuS method is introduced to solve this problem.
Seeing What You Miss: Vision-Language Pre-Training With Semantic Completion Learning
Yatai Ji (Tsinghua University), Wei Liu (Tencent)
RecognitionRetrievalTransformerVision Language ModelContrastive LearningImageVideoText
🎯 What it does: A semantic completion learning (SCL) task is proposed and implemented to enhance the global alignment capability between images/videos and text through cross-modal information completion.
Seeing What You Said: Talking Face Generation Guided by a Lip Reading Expert
Jiadong Wang (National University of Singapore), Haizhou Li (University of Electronic Science and Technology of China)
GenerationConvolutional Neural NetworkTransformerGenerative Adversarial NetworkContrastive LearningVideoAudio
🎯 What it does: In this study, the authors propose a speech-driven facial generation model called TalkLip, which significantly enhances the readability and lip-sync quality of the generated videos by using lip-reading experts to supervise the generation process.
Seeing With Sound: Long-range Acoustic Beamforming for Multimodal Scene Understanding
Praneeth Chakravarthula (Princeton University), Felix Heide (Princeton University)
Object DetectionAutonomous DrivingConvolutional Neural NetworkGenerative Adversarial NetworkMultimodalityAudio
🎯 What it does: A multimodal scene understanding method based on long-distance acoustic beamforming is proposed, and the resolution of acoustic signals is enhanced through neural aperture expansion.
SegLoc: Learning Segmentation-Based Representations for Privacy-Preserving Visual Localization
Maxime Pietrantoni (Czech Technical University in Prague), Gabriela Csurka (NAVER LABS Europe)
SegmentationRetrievalSafty and PrivacyConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes the SegLoc framework, which utilizes self-supervised fine-grained segmentation to generate sparse 3D maps for privacy-preserving visual localization.
Selective Structured State-Spaces for Long-Form Video Understanding
Jue Wang (Amazon Prime Video), Raffay Hamid (Amazon Prime Video)
RecognitionOptimizationComputational EfficiencyTransformerContrastive LearningVideo
🎯 What it does: The Selective S4 (S5) model is proposed, which adaptively selects important image tokens from videos using a lightweight mask generator, significantly improving performance in long-term video understanding tasks.
Self-Correctable and Adaptable Inference for Generalizable Human Pose Estimation
Zhehan Kan (Southern University of Science and Technology), Zhihai He (Southern University of Science and Technology)
Pose EstimationConvolutional Neural NetworkImage
🎯 What it does: A self-correcting adaptive inference (SCAI) framework is proposed, which utilizes a trained feedback network to evaluate prediction errors and corrects human pose estimation results in real-time through a correction network during testing, thereby enhancing the model's generalization performance in unknown scenarios.
Self-Guided Diffusion Models
Vincent Tao Hu (University of Amsterdam), Cees G. M. Snoek (University of Amsterdam)
SegmentationGenerationDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes an unsupervised self-guided diffusion model, which generates pseudo-labels, bounding boxes, and segmentation masks through self-supervised feature extraction and clustering, directly providing multi-scale guidance for the diffusion process, replacing traditional manual annotation guidance.
Self-Positioning Point-Based Transformer for Point Cloud Understanding
Jinyoung Park (Korea University), Hyunwoo J. Kim (Meta Reality Labs)
RecognitionSegmentationTransformerPoint Cloud
🎯 What it does: A Transformer framework named SPoTr is proposed for point cloud understanding, capable of capturing both local and global geometric contexts simultaneously.
Self-Supervised 3D Scene Flow Estimation Guided by Superpoints
Yaqi Shen (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)
Autonomous DrivingOptimizationRecurrent Neural NetworkOptical FlowPoint Cloud
🎯 What it does: An end-to-end self-supervised 3D scene flow estimation framework is proposed, utilizing dynamic superpoint generation and superpoint-guided flow refinement to achieve point-level flow prediction.
Self-Supervised AutoFlow
Hsin-Ping Huang (Google Research), Deqing Sun (Google Research)
Data SynthesisDomain AdaptationKnowledge DistillationSupervised Fine-TuningOptical FlowVideo
🎯 What it does: Using self-supervised loss as a search metric, AutoFlow is trained to generate synthetic training sets for the target domain, thereby learning optical flow from unlabeled real videos.
Self-Supervised Blind Motion Deblurring With Deep Expectation Maximization
Ji Li (National University of Singapore), Hui Ji (National University of Singapore)
RestorationConvolutional Neural NetworkImage
🎯 What it does: A data-free self-supervised deep learning method is proposed, utilizing Monte Carlo Expectation Maximization to train the network, recovering uniform and non-uniform motion blurred images caused by camera shake.
Self-Supervised Geometry-Aware Encoder for Style-Based 3D GAN Inversion
Yushi Lan (Nanyang Technological University), Bo Dai (Shanghai AI Laboratory)
GenerationData SynthesisNeural Radiance FieldGenerative Adversarial NetworkImage
🎯 What it does: A self-supervised learning-based 3D GAN inverse encoding framework called E3DGE is proposed, which can recover high-quality 3D shapes and textures using only a single 2D image of a face and achieve editable perspective-consistent generation.
Self-Supervised Image-to-Point Distillation via Semantically Tolerant Contrastive Loss
Anas Mahmoud (University of Toronto), Steven L. Waslander (University of Toronto)
SegmentationDomain AdaptationAutonomous DrivingRepresentation LearningContrastive LearningImagePoint Cloud
🎯 What it does: A semantic-tolerant contrastive loss and class-agnostic balanced loss are proposed for self-supervised image-to-point cloud representation learning.
Self-Supervised Implicit Glyph Attention for Text Recognition
Tongkun Guan (Shanghai Jiao Tong University), Wei Shen (Shanghai Jiao Tong University)
RecognitionSegmentationContrastive LearningImage
🎯 What it does: The paper proposes a self-supervised implicit stroke attention mechanism SIGA for scene text recognition.
Self-Supervised Learning for Multimodal Non-Rigid 3D Shape Matching
Dongliang Cao (University of Bonn), Florian Bernard (University of Bonn)
Representation LearningGraph Neural NetworkContrastive LearningMultimodalityPoint CloudMeshBiomedical Data
🎯 What it does: A self-supervised multimodal non-rigid 3D shape matching framework is proposed, capable of simultaneously handling triangular meshes, complete point clouds, and partial point clouds, achieving correspondences across different modalities.
Self-Supervised Learning From Images With a Joint-Embedding Predictive Architecture
Mahmoud Assran (Meta AI), Nicolas Ballas (Meta AI)
ClassificationRetrievalRepresentation LearningTransformerImage
🎯 What it does: This paper proposes a new self-supervised learning framework I-JEPA, which predicts feature representations of multiple target blocks in the same image using a single context block, without the need for manual data augmentation.
Self-Supervised Non-Uniform Kernel Estimation With Flow-Based Motion Prior for Blind Image Deblurring
Zhenxuan Fang (Xidian University), Guangming Shi (Xidian University)
RestorationConvolutional Neural NetworkFlow-based ModelImage
🎯 What it does: A self-supervised framework for non-uniform blur kernel estimation is proposed, utilizing a flow model to predict pixel-level blur kernels in the latent space, and embedding the estimated kernels into a deblurring network through a multi-scale kernel attention module.
Self-Supervised Pre-Training With Masked Shape Prediction for 3D Scene Understanding
Li Jiang (Max Planck Institute for Informatics), Bernt Schiele (Max Planck Institute for Informatics)
Object DetectionSegmentationRepresentation LearningTransformerContrastive LearningPoint Cloud
🎯 What it does: A self-supervised pre-training method based on Masked Shape Prediction (MSP) is proposed for learning general features from 3D scene point clouds.
Self-Supervised Representation Learning for CAD
Benjamin T. Jones (University of Washington), Adriana Schulz (University of Washington)
ClassificationSegmentationRepresentation LearningGraph Neural NetworkContrastive LearningMesh
🎯 What it does: Perform geometric self-supervised learning on CAD B-Rep to generate face-level embeddings and use these embeddings for few-shot segmentation and classification.
Self-Supervised Super-Plane for Neural 3D Reconstruction
Botao Ye (University of Chinese Academy of Sciences), Ming-Hsuan Yang (University of California Merced)
SegmentationDepth EstimationNeural Radiance FieldPoint Cloud
🎯 What it does: An unsupervised self-supervised hyperplane constraint is proposed to improve neural implicit surface reconstruction based on volumetric rendering, achieving smoother and more complete reconstructions in texture-sparse planar regions.
Self-Supervised Video Forensics by Audio-Visual Anomaly Detection
Chao Feng (University of Michigan), Andrew Owens (University of Michigan)
Anomaly DetectionTransformerVideoAudio
🎯 What it does: A video forensics method based on self-supervised anomaly detection is proposed, which identifies forged videos by learning the feature distribution of audio-video synchronization in real videos.
SelfME: Self-Supervised Motion Learning for Micro-Expression Recognition
Xinqi Fan (City University of Hong Kong), Hong Yan (City University of Hong Kong)
RecognitionTransformerContrastive LearningOptical FlowImageVideo
🎯 What it does: This paper proposes an end-to-end self-supervised motion learning framework called SelfME for micro-expression recognition, which first learns facial movements through a self-supervised reconstruction task and then classifies them using a symmetric contrast Vision Transformer.
Semantic Human Parsing via Scalable Semantic Transfer Over Multiple Label Domains
Jie Yang (Chinese University of Hong Kong), Ruimao Zhang (Tencent)
SegmentationConvolutional Neural NetworkTransformerImage
🎯 What it does: This study proposes a Scalable Semantic Transfer (SST) training framework for sharing semantic knowledge across multi-label domains, thereby enhancing the performance of human voxel segmentation networks.
Semantic Ray: Learning a Generalizable Semantic Field With Cross-Reprojection Attention
Fangfu Liu (Tsinghua University), Yueqi Duan (Tsinghua University)
SegmentationData SynthesisConvolutional Neural NetworkNeural Radiance FieldPoint Cloud
🎯 What it does: A semantic radiance field S-Ray that can generalize across scenes has been constructed, capable of directly rendering semantic maps in unseen scenes.
Semantic Scene Completion With Cleaner Self
Fengyun Wang (Nanjing University of Science and Technology), Qianru Sun (Singapore Management University)
SegmentationKnowledge DistillationConvolutional Neural NetworkPoint Cloud
🎯 What it does: Proposes the Cleaner Self framework, which uses TSDF-CAD for clean depth training through a teacher network and employs knowledge distillation to guide the student network, thereby improving performance on semantic scene completion tasks with noisy depth inputs.
Semantic-Conditional Diffusion Networks for Image Captioning
Jianjie Luo (Sun Yat-sen University), Tao Mei (HiDream.ai Inc.)
GenerationRetrievalTransformerDiffusion modelImageText
🎯 What it does: A Semantic Conditional Diffusion Network (SCD-Net) is proposed, which utilizes a diffusion model combined with the semantic prior of images to achieve non-autoregressive image captioning.