ICCV 2023 Papers — Page 16
IEEE/CVF International Conference on Computer Vision · 2156 papers
Quality-Agnostic Deepfake Detection with Intra-model Collaborative Learning
Binh M. Le (Sungkyunkwan University), Simon S. Woo (Sungkyunkwan University)
ClassificationAnomaly DetectionConvolutional Neural NetworkGenerative Adversarial NetworkVideo
🎯 What it does: A general quality-independent deepfake detection framework QAD is proposed, capable of simultaneously identifying deepfake videos of different qualities within a single model.
Query Refinement Transformer for 3D Instance Segmentation
Jiahao Lu (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
Object DetectionSegmentationTransformerContrastive LearningPoint Cloud
🎯 What it does: This paper proposes QueryFormer, a query-enhanced transformer for 3D instance segmentation, addressing the issues of query sampling and interference from noisy background queries.
Query6DoF: Learning Sparse Queries as Implicit Shape Prior for Category-Level 6DoF Pose Estimation
Ruiqi Wang (Huazhong University of Science and Technology), Wenyu Liu (Huazhong University of Science and Technology)
Pose EstimationPoint Cloud
🎯 What it does: A network named Query6DoF is proposed, utilizing sparse learnable queries as category-level shape priors, achieving query instantiation deformation and matching in the point cloud feature space through an attention mechanism, ultimately directly regressing the target's 6DoF pose and size.
R-Pred: Two-Stage Motion Prediction Via Tube-Query Attention-Based Trajectory Refinement
Sehwan Choi (Hanyang University), Jun Won Choi (Qualcomm)
Autonomous DrivingTransformerMultimodality
🎯 What it does: A two-stage trajectory prediction framework R-Pred is proposed, which first generates multimodal trajectory candidates using an initial network, and then refines each candidate through tube-query scene attention and proposal-level interaction attention.
R3D3: Dense 3D Reconstruction of Dynamic Scenes from Multiple Cameras
Aron Schmied (ETH Zurich), Fisher Yu (ETH Zurich)
Pose EstimationDepth EstimationAutonomous DrivingRecurrent Neural NetworkSimultaneous Localization and MappingPoint Cloud
🎯 What it does: A real-time dense 3D reconstruction and self-pose estimation system based on multiple cameras, R3D3, is proposed, capable of generating high-quality continuous point clouds in dynamic outdoor environments.
RANA: Relightable Articulated Neural Avatars
Umar Iqbal (NVIDIA), Jan Kautz (NVIDIA)
GenerationData SynthesisPose EstimationGenerative Adversarial NetworkImageVideo
🎯 What it does: A full-body neural portrait that can be rendered under any pose, viewpoint, and lighting can be learned using only short monocular RGB videos.
Random Boxes Are Open-world Object Detectors
Yanghao Wang (Nanyang Technological University), Hanwang Zhang (Nanyang Technological University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes RandBox, an open-world object detection framework that uses randomly generated boxes during training, capable of recognizing known categories while labeling unannotated objects as 'unknown'.
Random Sub-Samples Generation for Self-Supervised Real Image Denoising
Yizhong Pan (Sichuan University), Chao Ren (Sichuan University)
RestorationImageBenchmark
🎯 What it does: The SDAP framework is proposed, achieving unsupervised real image denoising through Random Subsample Generation (RSG) and Circular Sampling Differential Loss (CSDBSN).
Randomized Quantization: A Generic Augmentation for Data Agnostic Self-supervised Learning
Huimin Wu (Hong Kong University of Science and Technology), Zhirong Wu (Microsoft Research Asia)
Representation LearningData-Centric LearningContrastive LearningImageMultimodalityPoint CloudBenchmarkAudio
🎯 What it does: Randomized Quantization is proposed as a general data augmentation method for self-supervised representation learning; this method performs random non-uniform quantization on each channel along the channel dimension, randomly selecting quantization intervals and reconstruction values, thereby suppressing intra-interval information while preserving inter-interval information; this augmentation is combined with traditional contrastive learning frameworks (MoCo-v3, BYOL) to validate its cross-modal generality.
RankMatch: Fostering Confidence and Consistency in Learning with Noisy Labels
Ziyi Zhang (Nanjing University), Guanbin Li (Sun Yat-sen University)
ClassificationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A learning framework for noisy labels (LNL) named RankMatch is proposed, which integrates confidence voting sample selection and rank-based contrastive loss to enhance classification performance on noisy data.
RankMixup: Ranking-Based Mixup Training for Network Calibration
Jongyoun Noh (Yonsei University), Bumsub Ham (Yonsei University)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a hybrid training framework based on ranking relationships, called RankMixup, to improve the confidence calibration of deep networks.
Rapid Adaptation in Online Continual Learning: Are We Evaluating It Right?
Hasan Abed Al Kader Hammoud (King Abdullah University of Science and Technology), Bernard Ghanem (King Abdullah University of Science and Technology)
ClassificationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: Reassess the adaptability metrics of online continual learning algorithms and propose the 'Near-Future Accuracy' evaluation method.
Rapid Network Adaptation: Learning to Adapt Neural Networks Using Test-Time Feedback
Teresa Yeo (Swiss Federal Institute of Technology Lausanne), Amir Zamir (Swiss Federal Institute of Technology Lausanne)
SegmentationDepth EstimationDomain AdaptationReinforcement LearningImage
🎯 What it does: A closed-loop adaptive framework named Rapid Network Adaptation (RNA) is proposed, which quickly adjusts the frozen main network during inference by learning a small controller network that utilizes self-supervised signals.
RawHDR: High Dynamic Range Image Reconstruction from a Single Raw Image
Yunhao Zou (Hangzhou Dianzi University), Ying Fu (Beijing Institute of Technology)
RestorationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a HDR reconstruction framework called RawHDR based on a single raw image, which can directly generate a 20-bit HDR image from 14-bit raw data.
Ray Conditioning: Trading Photo-consistency for Photo-realism in Multi-view Image Generation
Eric Ming Chen (Cornell University), Abe Davis (Cornell University)
GenerationData SynthesisPose EstimationGenerative Adversarial NetworkImage
🎯 What it does: A 2D GAN method based on light conditions is proposed, achieving multi-view image generation and supporting explicit viewpoint control without requiring explicit 3D structure.
RbA: Segmenting Unknown Regions Rejected by All
Nazir Nayal (Koc University), Fatma Güney (Koc University)
SegmentationAnomaly DetectionTransformerSupervised Fine-TuningImage
🎯 What it does: A method for unknown object segmentation based on mask classification is proposed, and the detection of unknown pixels is achieved through the 'rejected by all known classes' (RbA) scoring function.
RCA-NOC: Relative Contrastive Alignment for Novel Object Captioning
Jiashuo Fan (Tsinghua University), Lei Zhang (International Digital Economy Academy)
Object DetectionGenerationTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a relative contrastive learning framework RCA-NOC, which utilizes visual semantic labels and their ranking information obtained from CLIP to achieve alignment between vision and language by contrasting positive and negative labels, thereby improving the accuracy of image captions for new objects.
Re-mine, Learn and Reason: Exploring the Cross-modal Semantic Correlations for Language-guided HOI detection
Yichao Cao (Southeast University), Chang Xu (University of Sydney)
Object DetectionKnowledge DistillationTransformerVision Language ModelImageTextMultimodality
🎯 What it does: A unified framework RmLR is proposed to enhance language-guided HOI detection through re-mining visual features, fine-grained sentence/word-level alignment, and knowledge distillation.
Re-ReND: Real-Time Rendering of NeRFs across Devices
Sara Rojas (King Abdullah University of Science and Technology), Bernard Ghanem (King Abdullah University of Science and Technology)
GenerationComputational EfficiencyNeural Radiance FieldMesh
🎯 What it does: Transform the pre-trained NeRF into a sparse grid and lightweight light field embedded table representation that can be rendered in real-time on resource-constrained devices such as mobile phones and AR/VR headsets, without calling MLP at all.
Re:PolyWorld - A Graph Neural Network for Polygonal Scene Parsing
Stefano Zorzi (Graz University of Technology), Friedrich Fraundorfer (Graz University of Technology)
Object DetectionSegmentationGraph Neural NetworkSupervised Fine-TuningImageGraph
🎯 What it does: This paper presents Re:PolyWorld, a graph neural network that generates polygons by extracting vertices from images and using differentiable optimal transport.
ReactioNet: Learning High-Order Facial Behavior from Universal Stimulus-Reaction by Dyadic Relation Reasoning
Xiaotian Li (State University of New York at Binghamton), Lijun Yin (State University of New York at Binghamton)
ClassificationRecognitionDomain AdaptationConvolutional Neural NetworkGraph Neural NetworkTransformerContrastive LearningVideoMultimodality
🎯 What it does: A large-scale synchronous stimulus-response pair database, ReactioNet, has been constructed, and a bidirectional relationship reasoning framework, DRR, has been proposed to enhance emotion recognition and facial action unit detection performance by utilizing the high-order relationships between stimulus scenarios and facial expressions.
Read-only Prompt Optimization for Vision-Language Few-shot Learning
Dongjun Lee (Korea University), Hyunwoo J. Kim (Korea University)
ClassificationDomain AdaptationOptimizationTransformerPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes a Read-only Prompt Optimization (RPO) method, which prevents unnecessary shifts in the internal representations of pre-trained vision-language models during fine-tuning through a masked attention mechanism and special token initialization, thereby achieving efficient and robust few-shot learning.
Real-Time Neural Rasterization for Large Scenes
Jeffrey Yunfan Liu (Waabi), Raquel Urtasun (University of Toronto)
Autonomous DrivingNeural Radiance FieldMesh
🎯 What it does: This paper presents NeuRas, a real-time large-scale scene novel view synthesis method that combines rasterization and neural textures.
RealGraph: A Multiview Dataset for 4D Real-world Context Graph Generation
Haozhe Lin (Tsinghua University), Lu Fang (Tsinghua University)
Object DetectionObject TrackingGenerationRecurrent Neural NetworkVideo
🎯 What it does: Proposed the 4D Scene Context Graph Generation (CGG) task and constructed the first multi-view RGB video dataset RealGraph, while providing the baseline model MCGNet;
Realistic Full-Body Tracking from Sparse Observations via Joint-Level Modeling
Xiaozheng Zheng (ByteDance Inc), Xiaojie Jin
Pose EstimationTransformerVideo
🎯 What it does: This paper proposes a two-stage joint hierarchical modeling framework that can drive a 3D full-body virtual avatar in real-time and realistically using only sparse observations from HMD and hand controllers.
REAP: A Large-Scale Realistic Adversarial Patch Benchmark
Nabeel Hingun (University of California Berkeley), David Wagner (University of California Berkeley)
Object DetectionAdversarial AttackImageBenchmark
🎯 What it does: This paper presents REAP, a benchmark for evaluating adversarial patch attacks in large-scale, real-world environments, covering over 14,000 traffic sign images and providing tools for geometric and lighting transformations.
Reconciling Object-Level and Global-Level Objectives for Long-Tail Detection
Shaoyu Zhang (Institute of Automation), Silong Peng (Beijing Visystem Co. Ltd)
Object DetectionSupervised Fine-TuningImage
🎯 What it does: This paper proposes a multi-task learning framework called ROG, which combines object-level classification and global ranking tasks to enhance long-tail detection performance.
Reconstructed Convolution Module Based Look-Up Tables for Efficient Image Super-Resolution
Guandu Liu (Tsinghua University), Bin Wang (Tsinghua University)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: A novel Reconfigurable Convolution (RC) module has been developed, decoupling channel and spatial computations, and achieving n×n convolution through n² 1D LUTs, significantly enhancing the receptive field and reducing storage in single image super-resolution.
Reconstructing Groups of People with Hypergraph Relational Reasoning
Buzhen Huang (Southeast University), Yangang Wang (Southeast University)
SegmentationPose EstimationGraph Neural NetworkImageMesh
🎯 What it does: This paper proposes a hypergraph relational reasoning network to recover multiple 3D human body meshes from a single large-scale crowd image.
Reconstructing Interacting Hands with Interaction Prior from Monocular Images
Binghui Zuo (Southeast University), Yangang Wang (Southeast University)
GenerationPose EstimationConvolutional Neural NetworkTransformerAuto EncoderImageMultimodality
🎯 What it does: A monocular interactive gesture reconstruction framework based on interactive priors and interactive adjacency heatmaps is proposed, utilizing VAE to generate latent space and sampling through ViT to fuse features;
Recovering a Molecule's 3D Dynamics from Liquid-phase Electron Microscopy Movies
Enze Ye (Peking University), He Sun (Peking University)
RestorationGenerationRecurrent Neural NetworkAuto EncoderVideoTime Series
🎯 What it does: The study investigates how to recover the three-dimensional dynamics of molecules from liquid-phase electron microscopy (liquid-phase EM) videos.
RecRecNet: Rectangling Rectified Wide-Angle Images by Thin-Plate Spline Model and DoF-based Curriculum Learning
Kang Liao (Beijing Jiaotong University), Yao Zhao (Beijing Jiaotong University)
Image TranslationRestorationConvolutional Neural NetworkImage
🎯 What it does: A new RecRecNet model is proposed to re-correct undistorted wide-angle images into rectangular boundaries while maintaining content integrity.
Recursive Video Lane Detection
Dongkwon Jin (Korea University), Chang-Su Kim (Korea University)
SegmentationAutonomous DrivingConvolutional Neural NetworkOptical FlowVideo
🎯 What it does: A recursive video lane detection algorithm RVLD has been developed, which utilizes single-frame historical information to achieve continuous video lane detection through motion estimation and feature reconstruction.
RecursiveDet: End-to-End Region-Based Recursive Object Detection
Jing Zhao (East China Normal University), Qingli Li (East China Normal University)
Object DetectionTransformerImage
🎯 What it does: A recursive decoder is proposed, utilizing parameter sharing and bounding box position encoding, significantly improving the performance of end-to-end region detectors while reducing the model parameter count.
RED-PSM: Regularization by Denoising of Partially Separable Models for Dynamic Imaging
Berk Iskender, Yoram Bresler
RestorationOptimizationSupervised Fine-TuningImageVideoMagnetic Resonance ImagingComputed Tomography
🎯 What it does: The RED-PSM method is proposed, combining low-rank partially separable models and denoising-based regularization to address the undersampling reconstruction problem in dynamic imaging.
Reducing Training Time in Cross-Silo Federated Learning Using Multigraph Topology
Tuong Do (AIOZ), Anh Nguyen (AIOZ)
OptimizationFederated LearningComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: A cross-silo federated learning topology based on multiple graphs is proposed, which utilizes multi-graph analysis to achieve different states and allows isolated nodes to skip aggregation in each communication round, significantly reducing training cycles.
Ref-NeuS: Ambiguity-Reduced Neural Implicit Surface Learning for Multi-View Reconstruction with Reflection
Wenhang Ge (Hong Kong University of Science and Technology), Ying-Cong Chen (Hong Kong University of Science and Technology)
Anomaly DetectionOptimizationNeural Radiance FieldImage
🎯 What it does: This paper studies the blurring problem caused by reflective surfaces in multi-view reconstruction and proposes the Ref-NeuS framework, which utilizes explicit reflection scores in photometric loss and reflection direction-related radiance fields to enhance the geometric and normal quality of reflective objects.
RefEgo: Referring Expression Comprehension Dataset from First-Person Perception of Ego4D
Shuhei Kurita (RIKEN), Eri Onami (Nara Institute of Science and Technology)
RecognitionObject DetectionObject TrackingTransformerVision Language ModelVideo
🎯 What it does: Created and released the RefEgo dataset, the first large-scale, real-world dataset for understanding video reference expressions based on Ego4D first-person videos.
Reference-guided Controllable Inpainting of Neural Radiance Fields
Ashkan Mirzaei (Samsung AI Centre), Igor Gilitschenski (University of Toronto)
RestorationGenerationDiffusion modelNeural Radiance FieldImage
🎯 What it does: Utilizing a single patched view for 3D completion of NeRF scenes, achieving perspective-consistent area filling.
Referring Image Segmentation Using Text Supervision
Fang Liu (Dalian University of Technology), Rynson Lau (City University of Hong Kong)
Object DetectionSegmentationTransformerContrastive LearningImageText
🎯 What it does: A weakly supervised image segmentation framework is proposed that uses only text expressions as supervision, optimizing the positioning of targets through text-image response and generating pseudo-labels to train the segmentation network.
ReFit: Recurrent Fitting Network for 3D Human Recovery
Yufu Wang (University of Pennsylvania), Kostas Daniilidis (University of Pennsylvania)
Pose EstimationRecurrent Neural NetworkImage
🎯 What it does: A cyclic fitting network named ReFit is proposed, which recovers 3D human posture and shape from a single image through an iterative feedback-update loop.
ReGen: A good Generative Zero-Shot Video Classifier Should be Rewarded
Adrian Bulat (Samsung AI Cambridge), Georgios Tzimiropoulos (Samsung AI Cambridge)
ClassificationRecognitionGenerationReinforcement LearningVideo
🎯 What it does: A generative video captioning model is trained through reinforcement learning, enabling its output captions to be used for zero-shot/few-shot video action classification.
RegFormer: An Efficient Projection-Aware Transformer Network for Large-Scale Point Cloud Registration
Jiuming Liu (Shanghai Jiao Tong University), Hesheng Wang (Shanghai Jiao Tong University)
Autonomous DrivingOptimizationTransformerFlow-based ModelPoint Cloud
🎯 What it does: An end-to-end RegFormer network is proposed for large-scale point cloud registration, eliminating the dependence on keypoint detection, feature description, and RANSAC post-processing, and capable of directly estimating rigid transformations from raw LiDAR point clouds.
Regularized Mask Tuning: Uncovering Hidden Knowledge in Pre-Trained Vision-Language Models
Kecheng Zheng (Zhejiang University), Yujun Shen (Ant Group)
ClassificationTransformerVision Language ModelImage
🎯 What it does: For pre-trained vision-language models, a method is proposed to regularize mask fine-tuning by freezing model parameters through learnable binary masks.
Regularized Primitive Graph Learning for Unified Vector Mapping
Lei Wang (Huawei Technologies), Jingwei Huang (Huawei Technologies)
Object DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: A unified framework called GraphMapper is proposed, which transforms various vector map tasks into primitive graph reconstruction, achieving end-to-end extraction of buildings and roads.
Rehearsal-Free Domain Continual Face Anti-Spoofing: Generalize More and Forget Less
Rizhao Cai (Nanyang Technological University), Alex Kot (Nanyang Technological University)
RecognitionDomain AdaptationTransformerContrastive LearningImage
🎯 What it does: This study investigates a domain continual learning (DCL) facial spoof detection (FAS) model under conditions of no replay buffer and low sample size, and proposes a new replay-free training framework.
Reinforce Data, Multiply Impact: Improved Model Accuracy and Robustness with Dataset Reinforcement
Fartash Faghri (Apple), Oncel Tuzel (Apple)
ClassificationObject DetectionSegmentationKnowledge DistillationData-Centric LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: By precomputing and storing the outputs of a strong teacher model under various data augmentations on the training set, a reinforced dataset (such as ImageNet+) is constructed to improve the accuracy and robustness of any model without additional training costs.
Reinforced Disentanglement for Face Swapping without Skip Connection
Xiaohang Ren (Xiaobing.AI), Baoyuan Wang
Image TranslationGenerationRetrievalGenerative Adversarial NetworkImageVideo
🎯 What it does: This paper proposes a facial swapping framework without skip connections and dual encoders (FNID and NFA), which can completely remove target identity information while preserving non-identity attributes of the target.
ReLeaPS : Reinforcement Learning-based Illumination Planning for Generalized Photometric Stereo
Jun Hoong Chan (Peking University), Boxin Shi (Peking University)
OptimizationReinforcement LearningImage
🎯 What it does: A reinforcement learning-based online lighting planning method called ReLeaPS is proposed for surface normal estimation using general photometric stereo with a limited number of light sources.
Relightify: Relightable 3D Faces from a Single Image via Diffusion Models
Foivos Paraperas Papantoniou (Imperial College London), Stefanos Zafeiriou (Imperial College London)
RestorationGenerationDiffusion modelImage
🎯 What it does: Using a single facial image, this paper combines an unsupervised diffusion model to simultaneously recover UV textures and BRDF (diffuse reflection, specular reflection, normals), generating a 3D facial model that can be rendered under arbitrary lighting.
Remembering Normality: Memory-guided Knowledge Distillation for Unsupervised Anomaly Detection
Zhihao Gu (Shanghai Jiao Tong University), Lizhuang Ma (CATL)
Anomaly DetectionKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: Proposes the Memory-guided Knowledge Distillation (MemKD) framework, which utilizes a key-value structured Normality Recall Memory (NR Memory) and Normality Embedding Learning (NEL) to enhance the normal feature representation of the student network, thereby addressing the 'normality forgetting' problem in traditional KD methods.
ReMoDiffuse: Retrieval-Augmented Motion Diffusion Model
Mingyuan Zhang (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
GenerationRetrievalTransformerDiffusion modelTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes a retrieval-enhanced 3D human motion diffusion model, ReMoDiffuse, for text-driven motion generation.
Removing Anomalies as Noises for Industrial Defect Localization
Fanbin Lu (Chinese University of Hong Kong), Jiaya Jia (Chinese University of Hong Kong)
Anomaly DetectionDiffusion modelImageBenchmark
🎯 What it does: This paper proposes a denoising-based anomaly detection method using diffusion models, which adds noise to the input images and utilizes the KL divergence of the intermediate steps in the diffusion model as pixel-level anomaly scores, while also combining feature reconstruction errors to achieve precise anomaly localization and high-quality reconstruction.
RenderIH: A Large-Scale Synthetic Dataset for 3D Interacting Hand Pose Estimation
Lijun Li (Alibaba Group), Chen Chen (University of Central Florida)
Data SynthesisPose EstimationTransformerImage
🎯 What it does: This paper proposes a large-scale synthetic dataset RenderIH to enhance 3D hand pose estimation under single RGB images, and conducts experimental validation based on the Transformer-based TransHand network.
Rendering Humans from Object-Occluded Monocular Videos
Tiange Xiang (Stanford University), Li Fei-Fei (Stanford University)
RestorationGenerationPose EstimationNeural Radiance FieldVideo
🎯 What it does: This paper proposes a neural rendering framework called OccNeRF for reconstructing complete 3D human models from monocular videos with occluded objects.
ReNeRF: Relightable Neural Radiance Fields with Nearfield Lighting
Yingyan Xu (ETH Zurich), Paulo Gotardo (ETH Zurich)
GenerationNeural Radiance FieldImage
🎯 What it does: A re-illuminable neural radiance field (ReNeRF) is proposed, which is trained under conditions of capturing with only a few local light sources, achieving complete control over viewpoint and illumination.
Replay: Multi-modal Multi-view Acted Videos for Casual Holography
Roman Shapovalov (Meta), Natalia Neverova (Meta)
SegmentationData SynthesisNeural Radiance FieldVideoMultimodalityBenchmark
🎯 What it does: This paper presents the Replay dataset, which includes multi-view, multi-modal social scene videos, and establishes two new view synthesis benchmarks (flyaround and acting) to evaluate various NeRF-based methods.
RepQ-ViT: Scale Reparameterization for Post-Training Quantization of Vision Transformers
Zhikai Li (Institute of Automation, Chinese Academy of Sciences), Qingyi Gu (Institute of Automation, Chinese Academy of Sciences)
Object DetectionSegmentationCompressionTransformerImage
🎯 What it does: A new post-training quantization framework RepQ-ViT is proposed to compress Vision Transformer models.
Representation Disparity-aware Distillation for 3D Object Detection
Yanjing Li (Beihang University), Xianbin Cao (Beihang University)
Object DetectionKnowledge DistillationPoint Cloud
🎯 What it does: Proposes a representation difference-aware distillation (RDD) method based on information bottleneck to enhance the performance of ultra-small 3D detectors.
Representation Uncertainty in Self-Supervised Learning as Variational Inference
Hiroki Nakamura (Panasonic Holdings Corporation), Tadahiro Taniguchi (Ritsumeikan University)
Representation LearningContrastive LearningImage
🎯 What it does: The research views self-supervised learning as variational inference and proposes the VI-SimSiam method, which can simultaneously learn representations and their uncertainties.
Residual Pattern Learning for Pixel-Wise Out-of-Distribution Detection in Semantic Segmentation
Yuyuan Liu (Australian Institute for Machine Learning), Gustavo Carneiro (University of Surrey)
SegmentationAnomaly DetectionConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A Residual Pattern Learning (RPL) module and Context Robust Contrastive Learning (CoroCL) have been designed to achieve pixel-level Out-of-Distribution (OoD) detection on a frozen semantic segmentation network while maintaining segmentation accuracy.
ResQ: Residual Quantization for Video Perception
Davide Abati (Qualcomm AI Research), Amirhossein Habibian (Qualcomm AI Research)
SegmentationPose EstimationCompressionComputational EfficiencyConvolutional Neural NetworkVideo
🎯 What it does: A low-bit quantization method based on video frame residuals, ResQ, is proposed, which utilizes the low variance characteristics of residuals to perform low-bit quantization on subsequent frames while maintaining high accuracy for key frames, thereby accelerating video perception tasks.
ReST: A Reconfigurable Spatial-Temporal Graph Model for Multi-Camera Multi-Object Tracking
Cheng-Che Cheng (National Tsing Hua University), Shang-Hong Lai (National Tsing Hua University)
Object TrackingGraph Neural NetworkVideo
🎯 What it does: Multi-camera multi-target tracking based on a reconfigurable space-time graph model
Rethinking Amodal Video Segmentation from Learning Supervised Signals with Object-centric Representation
Ke Fan (Fudan University), Yanwei Fu (Fudan University)
Object DetectionSegmentationAutonomous DrivingConvolutional Neural NetworkVideo
🎯 What it does: Proposes the EoRaS method, which utilizes supervised visible masks and multi-view information for video intangible segmentation, achieving joint reasoning of shape and viewpoint priors through a BEV translation network and a multi-view fusion layer.
Rethinking Data Distillation: Do Not Overlook Calibration
Dongyao Zhu (University of California San Diego), Dongkuan Xu (North Carolina State University)
Knowledge DistillationImage
🎯 What it does: This study investigates the calibration problem of networks trained through data distillation (DDNN) and proposes two methods, Masked Temperature Scaling (MTS) and Masked Distillation Training (MDT), to address it.
Rethinking Fast Fourier Convolution in Image Inpainting
Tianyi Chu (Zhejiang University), Dongming Lu (Zhejiang University)
RestorationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a new unbiased fast Fourier convolution (UFFC) module for image inpainting tasks, which can avoid the spectral shift and spatial activation issues of traditional FFC in the frequency domain while ensuring a global receptive field.
Rethinking Mobile Block for Efficient Attention-based Models
Jiangning Zhang (Youtu Lab Tencent), Chengjie Wang (Youtu Lab Tencent)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerImage
🎯 What it does: A unified lightweight basic block called Meta Mobile Block (MMB) is proposed, from which an improved Inverted Residual Mobile Block (iRMB) is derived. Based on this, an efficient model EMO, consisting solely of iRMB, is constructed for dense prediction tasks such as image classification, object detection, and semantic segmentation.
Rethinking Multi-Contrast MRI Super-Resolution: Rectangle-Window Cross-Attention Transformer and Arbitrary-Scale Upsampling
Guangyuan Li (Zhejiang University), Wei Xing (Zhejiang University)
RestorationSuper ResolutionTransformerImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A new multi-contrast MRI super-resolution network McASSR is proposed to address the lack of direct interaction in existing methods with fixed integer upsampling ratios and windows.
Rethinking Point Cloud Registration as Masking and Reconstruction
Guangyan Chen (Beijing Institute of Technology), Yufeng Yue (Beijing Institute of Technology)
Object DetectionTransformerPoint Cloud
🎯 What it does: View point cloud registration as a masking and reconstruction task, proposing the Mask Reconstruction Auxiliary Network (MRA) to assist the main network in learning fine-grained geometry and overall structure, and based on this, designing the Mask Reconstruction Transformer (MRT) to achieve efficient and accurate registration.
Rethinking Pose Estimation in Crowds: Overcoming the Detection Information Bottleneck and Ambiguity
Mu Zhou, Alexander Mathis (Ecole Polytechnique Federale de Lausanne)
Pose EstimationConvolutional Neural NetworkTransformerImage
🎯 What it does: A two-stage pose estimation framework that combines low-level detection with conditional top-level processing (BUCTD) is proposed, utilizing a low-level pose detector to generate pose hints as conditional inputs for the top network.
Rethinking Range View Representation for LiDAR Segmentation
Lingdong Kong (Shanghai AI Laboratory), Ziwei Liu (Nanyang Technological University)
SegmentationAutonomous DrivingTransformerPoint Cloud
🎯 What it does: Proposes the RangeFormer Transformer framework and the STR scalable training strategy for semantic and panoramic segmentation of LiDAR point clouds in range view.
Rethinking Safe Semi-supervised Learning: Transferring the Open-set Problem to A Close-set One
Qiankun Ma (Sichuan University), Yan Wang (Sichuan University)
ClassificationContrastive LearningImage
🎯 What it does: This paper proposes a secure semi-supervised learning framework based on prototypical networks, transforming the open set problem into a closed set problem, and classifying on K+1 classes; it also introduces Iterative Negative Learning (INL) during the pseudo-labeling phase to leverage low-confidence samples.
Rethinking the Role of Pre-Trained Networks in Source-Free Domain Adaptation
Wenyu Zhang (Agency for Science Technology and Research), Chuan-Sheng Foo (Agency for Science Technology and Research)
Domain AdaptationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes to retain the pre-trained network (such as ImageNet weights) after training the source model and involve it in target domain adaptation, enhancing the quality of pseudo-labels through a co-learning strategy, thereby improving the performance of the source model in the target domain.
Rethinking Video Frame Interpolation from Shutter Mode Induced Degradation
Xiang Ji (University of Tokyo), Yinqiang Zheng (National Institute of Informatics)
RestorationGenerationConvolutional Neural NetworkOptical FlowVideoBenchmark
🎯 What it does: This paper presents the first video frame interpolation dataset RD-VFI affected by camera shutter modes (global shutter blur, rolling shutter distortion, rolling shutter + global reset) in real scenarios, and proposes a unified Progressive Mutual Boosting Network (PMBNet) to achieve intermediate frame interpolation under different degradation modes.
Rethinking Vision Transformers for MobileNet Size and Speed
Yanyu Li (Snap Inc.), Jian Ren (Snap Inc.)
ClassificationObject DetectionSegmentationComputational EfficiencyTransformerImage
🎯 What it does: This paper proposes a lightweight visual Transformer architecture named EfficientFormerV2, designed for efficient inference on mobile devices.
Retinexformer: One-stage Retinex-based Transformer for Low-light Image Enhancement
Yuanhao Cai (Tsinghua University), Yulun Zhang (ETH Zurich)
RestorationObject DetectionTransformerImage
🎯 What it does: A one-stage low-light image enhancement method based on the Retinex theory and Transformer (Retinexformer) is proposed.
Retro-FPN: Retrospective Feature Pyramid Network for Point Cloud Semantic Segmentation
Peng Xiang (Tsinghua University), Zhizhong Han (Wayne State University)
SegmentationTransformerPoint Cloud
🎯 What it does: Proposed Retro-FPN, which explicitly and recursively predicts and refines semantic features at each point in the encoder-decoder feature pyramid.
Revisit PCA-based Technique for Out-of-Distribution Detection
Xiaoyuan Guan (Sun Yat-sen University), Ruixuan Wang (Sun Yat-sen University)
Anomaly DetectionAuto EncoderImage
🎯 What it does: A post-processing method that integrates the regularized PCA reconstruction error with energy scores is proposed to improve OOD detection in deep learning models.
Revisiting Domain-Adaptive 3D Object Detection by Reliable, Diverse and Class-balanced Pseudo-Labeling
Zhuoxiao Chen (University of Queensland), Zi Huang (University of Queensland)
Object DetectionDomain AdaptationAutonomous DrivingPoint Cloud
🎯 What it does: This paper proposes an unsupervised domain adaptation framework for 3D object detection called ReDB, which utilizes reliable, diverse, and class-balanced pseudo-labels to achieve multi-class self-training.
Revisiting Foreground and Background Separation in Weakly-supervised Temporal Action Localization: A Clustering-based Approach
Qinying Liu (University of Science and Technology of China), Yixin Zhang (Institute of Artificial Intelligence, Hefei Comprehensive National Science Center)
RecognitionObject DetectionContrastive LearningVideo
🎯 What it does: A weakly supervised temporal action localization framework CASE based on unsupervised clustering is proposed, specifically addressing the foreground/background separation problem.
Revisiting Scene Text Recognition: A Data Perspective
Qing Jiang (South China University of Technology), Lianwen Jin (South China University of Technology)
RecognitionData-Centric LearningTransformerSupervised Fine-TuningImageBenchmark
🎯 What it does: This paper re-evaluates scene text recognition (STR) from a data perspective, finding that commonly used benchmarks have reached saturation. It constructs a large real dataset, Union14M, and proposes a challenge-driven benchmark.
Revisiting the Parameter Efficiency of Adapters from the Perspective of Precision Redundancy
Shibo Jie (Peking University), Zhi-Hong Deng (Peking University)
ClassificationOptimizationComputational EfficiencyTransformerImage
🎯 What it does: This paper proposes a low-precision adapter training method that allows for efficient fine-tuning in visual tasks with minimal storage requirements.
Revisiting Vision Transformer from the View of Path Ensemble
Shuning Chang (National University of Singapore), Mike Zheng Shou (Alibaba Group)
ClassificationComputational EfficiencyKnowledge DistillationTransformerImage
🎯 What it does: A new perspective on Vision Transformer (ViT) is proposed: viewing it as a collection of multiple paths of different lengths, and optimizing path combinations through techniques such as path pruning, EnsembleScale, and self-distillation to enhance ViT performance and support deeper, more efficient models.
RFD-ECNet: Extreme Underwater Image Compression with Reference to Feature Dictionary
Mengyao Li (Shanghai University), Zheyin Wang (Shanghai University)
CompressionAuto EncoderImage
🎯 What it does: A limit underwater image compression network RFD-ECNet based on an underwater multi-scale feature dictionary is designed, which can remove redundancy between different underwater images through a reference dictionary.
RFLA: A Stealthy Reflected Light Adversarial Attack in the Physical World
Donghua Wang (Zhejiang University), Xiaoqian Chen (Chinese Academy of Military Science)
Autonomous DrivingAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a physical adversarial attack based on reflected light (RFLA), which generates adjustable geometric shapes and colors of reflected light to deceive DNN models by using specular reflection of sunlight or flashlights, combined with colored transparent plastic sheets and paper cutouts.
Rickrolling the Artist: Injecting Backdoors into Text Encoders for Text-to-Image Synthesis
Lukas Struppek (Technical University of Darmstadt), Kristian Kersting (Technical University of Darmstadt)
GenerationData SynthesisAdversarial AttackTransformerSupervised Fine-TuningDiffusion modelImageText
🎯 What it does: This study investigates backdoor attacks on text-to-image synthesis models, demonstrating how to inject hidden triggers into pre-trained text encoders to control the model's generation of specified content or the addition of attributes.
RICO: Regularizing the Unobservable for Indoor Compositional Reconstruction
Zizhang Li (Zhejiang University), Yong Liu (Zhejiang University)
SegmentationGenerationNeural Radiance FieldPoint CloudMesh
🎯 What it does: The RICO method is proposed, achieving object-level decomposition and reconstruction by regularizing the invisible regions in indoor scenes.
RIGID: Recurrent GAN Inversion and Editing of Real Face Videos
Yangyang Xu (University of Hong Kong), Ping Luo (University of Hong Kong)
GenerationData SynthesisRecurrent Neural NetworkGenerative Adversarial NetworkOptical FlowVideo
🎯 What it does: A recursive framework called RIGID is proposed, which can simultaneously achieve GAN inversion and editing of real human face videos while maintaining spatiotemporal consistency.
RLIPv2: Fast Scaling of Relational Language-Image Pre-Training
Hangjie Yuan (Zhejiang University), Deli Zhao (Alibaba Group)
RecognitionObject DetectionTransformerVision Language ModelImageTextMultimodality
🎯 What it does: The RLIPv2 model is proposed, which combines the rapidly converging Asymmetric Language-Image Fusion (ALIF) and large-scale pseudo-labeled scene graph data to achieve large-scale relational language-image pre-training.
RLSAC: Reinforcement Learning Enhanced Sample Consensus for End-to-End Robust Estimation
Chang Nie (Shanghai Jiao Tong University), Hesheng Wang (Shanghai Jiao Tong University)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This work proposes RLSAC, a reinforcement learning-based sampling consensus framework that achieves end-to-end robust model estimation.
RMP-Loss: Regularizing Membrane Potential Distribution for Spiking Neural Networks
Yufei Guo (Intelligent Science and Technology Academy of CASIC), Zhe Ma (Intelligent Science and Technology Academy of CASIC)
ClassificationSpiking Neural NetworkImage
🎯 What it does: This paper proposes a regularization loss called RMP-Loss, which is used to directly reduce the membrane potential quantization error in SNNs, thereby improving the classification performance of SNNs.
Robo3D: Towards Robust and Reliable 3D Perception against Corruptions
Lingdong Kong (Shanghai AI Laboratory), Ziwei Liu (Nanyang Technological University)
Object DetectionSegmentationAutonomous DrivingPoint CloudBenchmark
🎯 What it does: Established the Robo3D benchmark, defined eight types of real-world LiDAR noise and faults, and systematically evaluated the robustness of 34 3D detection and segmentation models under different severity levels; also proposed two techniques, density-sensitive training framework and variable voxelization, to enhance robustness.
Robust e-NeRF: NeRF from Sparse & Noisy Events under Non-Uniform Motion
Weng Fei Low (National University of Singapore), Gim Hee Lee (National University of Singapore)
GenerationData SynthesisOptimizationNeural Radiance FieldImageVideo
🎯 What it does: This paper proposes a method to directly and robustly reconstruct NeRF from sparse, noisy event streams, specifically targeting moving event cameras under non-uniform motion.
Robust Evaluation of Diffusion-Based Adversarial Purification
Minjong Lee (POSTECH), Dongwoo Kim (POSTECH)
Adversarial AttackConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Evaluate the robustness of diffusion-based adversarial purification and propose more rigorous evaluation criteria and a multi-step purification strategy with gradual noise scheduling.
Robust Frame-to-Frame Camera Rotation Estimation in Crowded Scenes
Fabien Delattre (University of Massachusetts Amherst), Erik Learned-Miller (University of Massachusetts Amherst)
Pose EstimationComputational EfficiencyOptical FlowImageVideo
🎯 What it does: A robust camera rotation estimation method based on optical flow is proposed, utilizing Hough transform voting on SO(3) to find the most compatible rotation;
Robust Geometry-Preserving Depth Estimation Using Differentiable Rendering
Chi Zhang (Tencent), Chunhua Shen (Zhejiang University)
Depth EstimationImage
🎯 What it does: This paper proposes a geometry-preserving monocular depth estimation framework based on differentiable rendering, capable of achieving zero-shot cross-scene inference through training on mixed datasets without the use of additional 3D data or complete depth annotations.
Robust Heterogeneous Federated Learning under Data Corruption
Xiuwen Fang (Wuhan University), Xiyuan Yang (Wuhan University)
Federated LearningKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A robust training framework called AugHFL is proposed to address the issue of data corruption in heterogeneous federated learning, which can simultaneously suppress the negative effects of internal and external data corruption during both local training and global collaborative learning phases.
Robust Mixture-of-Expert Training for Convolutional Neural Networks
Yihua Zhang (Michigan State University), Sijia Liu (IBM Research)
OptimizationComputational EfficiencyAdversarial AttackConvolutional Neural NetworkMixture of ExpertsImage
🎯 What it does: A new robust mixture of experts (MoE-CNN) training framework is proposed and implemented, which enhances adversarial robustness while maintaining the efficiency of sparse activation in CNNs.
Robust Monocular Depth Estimation under Challenging Conditions
Stefano Gasperini (Technical University of Munich), Federico Tombari (Google)
Depth EstimationAutonomous DrivingKnowledge DistillationGenerative Adversarial NetworkImage
🎯 What it does: A general method named md4all is proposed, which enables monocular depth estimation models to maintain robustness under both ideal and adverse lighting and weather conditions (such as nighttime and rainy weather);
Robust Object Modeling for Visual Tracking
Yidong Cai (Nanjing University), Gangshan Wu (Nanjing University)
Object TrackingTransformerVideo
🎯 What it does: This paper proposes a robust object modeling framework named ROMTrack, which utilizes inherent templates, hybrid templates, and variation tokens to jointly learn target features, achieving end-to-end visual tracking.