ICCV 2023 Papers — Page 15
IEEE/CVF International Conference on Computer Vision · 2156 papers
PATMAT: Person Aware Tuning of Mask-Aware Transformer for Face Inpainting
Saman Motamed (Carnegie Mellon University), Fernando De la Torre (Carnegie Mellon University)
RestorationGenerationTransformerGenerative Adversarial NetworkImage
🎯 What it does: A facial restoration method called PATMAT has been developed, which utilizes a small number of reference images to personalize the Mask-Aware Transformer (MAT) for identity-preserving restoration of large facial gaps.
PC-Adapter: Topology-Aware Adapter for Efficient Domain Adaption on Point Clouds with Rectified Pseudo-label
Joonhyung Park (KAIST), Eunho Yang (KAIST)
Domain AdaptationGraph Neural NetworkPoint Cloud
🎯 What it does: A lightweight Adapter-based unsupervised domain adaptation framework for point clouds, called PC-Adapter, is proposed.
PDiscoNet: Semantically consistent part discovery for fine-grained recognition
Robert van der Klis (Wageningen University and Research), Diego Marcos (Inria)
ClassificationRecognitionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes PDiscoNet, which learns interpretable object parts by utilizing image-level labels in fine-grained classification tasks, treating parts as information bottlenecks for classification.
PEANUT: Predicting and Navigating to Unseen Targets
Albert J. Zhai (University of Illinois at Urbana-Champaign), Shenlong Wang (University of Illinois at Urbana-Champaign)
Object DetectionRobotic IntelligencePoint Cloud
🎯 What it does: A framework named PEANUT is proposed for ObjectGoal navigation based on explicit prediction, which predicts the occurrence probability of target objects in unexplored areas using a complete global semantic map, and selects long-term goals in a distance-weighted manner;
Perceptual Artifacts Localization for Image Synthesis Tasks
Lingzhi Zhang (University of Pennsylvania), Jianbo Shi (University of Pennsylvania)
RestorationSegmentationData SynthesisTransformerDiffusion modelImage
🎯 What it does: This paper constructs a synthetic dataset of 10,168 images with pixel-level defect annotations and trains a unified semantic segmentation model to locate perceptual defects; it also proposes a scalable repair pipeline for automatic defect repair and demonstrates its application in multiple tasks.
Perceptual Grouping in Contrastive Vision-Language Models
Kanchana Ranasinghe (Apple), Jonathon Shlens (Apple)
ClassificationSegmentationTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper studies the lack of perception of spatial location information of objects in contrastive vision-language models and proposes improvements through modifications in aggregation methods, pre-training strategies, and token subsampling to achieve perceptual grouping.
Periodically Exchange Teacher-Student for Source-Free Object Detection
Qipeng Liu (Fuzhou University), Zhifeng Yang (Fuzhou University)
Object DetectionDomain AdaptationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A periodic exchange teacher-student framework (PETS) is proposed for source-free object detection (SFOD), addressing the stability issues caused by the collapse of the teacher model in traditional Mean Teacher training.
Perpetual Humanoid Control for Real-time Simulated Avatars
Zhengyi Luo (Reality Labs Research), Weipeng Xu (Reality Labs Research)
Robotic IntelligenceReinforcement LearningDiffusion modelVideo
🎯 What it does: A physics-driven human controller (PHC) is proposed, capable of achieving high-fidelity motion imitation without external force assistance and naturally recovering in the event of a fall or input noise.
Persistent-Transient Duality: A Multi-Mechanism Approach for Modeling Human-Object Interaction
Hung Tran (Deakin University), Truyen Tran (Deakin University)
Pose EstimationRecurrent Neural NetworkGraph Neural NetworkVideo
🎯 What it does: A Persistent-Transient Duality model is proposed to predict human motion trajectories and object positions in human-object interaction (HOI).
Person Re-Identification without Identification via Event anonymization
Shafiq Ahmad (Istituto Italiano di Tecnologia), Alessio Del Bue (Istituto Italiano di Tecnologia)
RecognitionRetrievalSafty and PrivacyConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: A privacy-preserving portrait re-identification framework based on event cameras is proposed, utilizing a learning-based event stream anonymization to suppress image reconstruction attacks while maintaining re-identification performance.
Personalized Image Generation for Color Vision Deficiency Population
Shuyi Jiang (University of Sydney), Chang Xu (University of Sydney)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: A GAN based on triple latent variables has been designed and implemented, capable of end-to-end generating images that meet the needs of the color-blind community, while supporting varying degrees of personalization.
Personalized Semantics Excitation for Federated Image Classification
Haifeng Xia (Tulane University), Zhengming Ding (Tulane University)
ClassificationFederated LearningConvolutional Neural NetworkImageMultimodality
🎯 What it does: A mechanism called Personalized Semantic Incentive (PSE) is proposed to generate more accurate and locally adaptive image classification models for each client within the federated learning framework.
PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images
Yingfei Liu (MEGVII Technology), Xiangyu Zhang (MEGVII Technology)
Object DetectionSegmentationAutonomous DrivingTransformerImagePoint Cloud
🎯 What it does: Proposes the PETRv2 framework, which integrates multi-camera images to achieve 3D perception, extending PETR by incorporating temporal modeling and multi-task learning (3D object detection, BEV segmentation, 3D lane detection).
PG-RCNN: Semantic Surface Point Generation for 3D Object Detection
Inyong Koo (KAIST), Changick Kim (KAIST)
Object DetectionAutonomous DrivingTransformerPoint Cloud
🎯 What it does: This paper proposes PG-RCNN, a two-stage LiDAR 3D object detector that enhances detection accuracy by completing sparse point clouds through generating semantic surface points within each candidate box.
PGFed: Personalize Each Client's Global Objective for Federated Learning
Jun Luo (University of Pittsburgh), Shandong Wu (University of Pittsburgh)
Federated LearningImage
🎯 What it does: A new personalized federated learning framework called PGFed is proposed, allowing each client to optimize its global objective by explicitly aggregating local and global empirical risks, and based on this, an accelerated version called PGFedMo is introduced.
PhaseMP: Robust 3D Pose Estimation via Phase-conditioned Human Motion Prior
Mingyi Shi (University of Hong Kong), Jungdam Won (Seoul National University)
Pose EstimationAuto EncoderVideo
🎯 What it does: By introducing a phase-conditioned prior for human motion, frequency domain phase features are extracted using a periodic autoencoder, combined with a conditional VAE for 3D human pose estimation, and the prior is utilized during the optimization phase to achieve robust pose recovery and video-to-motion conversion.
Phasic Content Fusing Diffusion Model with Directional Distribution Consistency for Few-Shot Model Adaption
Teng Hu (Shanghai Jiao Tong University), Lizhuang Ma (Shanghai Jiao Tong University)
GenerationDomain AdaptationDiffusion modelImage
🎯 What it does: A few-shot diffusion model based on staged content fusion is proposed, incorporating directional distribution consistency loss and iterative cross-domain structure guidance strategy to achieve few-shot image generation and domain adaptation.
PHRIT: Parametric Hand Representation with Implicit Template
Zhisheng Huang (Wuhan University), Zhigang Tu (Wuhan University)
GenerationPose EstimationPoint CloudMesh
🎯 What it does: This paper presents PHRIT, a hand model that combines parameter grids with implicit templates, capable of achieving high-resolution and differentiable hand reconstruction based on skeletal and shape latent codes.
PhysDiff: Physics-Guided Human Motion Diffusion Model
Ye Yuan (NVIDIA), Jan Kautz (NVIDIA)
GenerationPose EstimationRobotic IntelligenceReinforcement LearningDiffusion modelVideoTextPhysics Related
🎯 What it does: A physics-guided motion diffusion model called PhysDiff is proposed, which embeds physical constraints during the diffusion process to generate physically feasible 3D human motion.
Physically-Plausible Illumination Distribution Estimation
Egor Ershov (Institute for Information Transmission Problems Russian Academy of Sciences), Michael S. Brown (York University)
Convolutional Neural NetworkImage
🎯 What it does: A global white balance method based on light distribution is proposed and implemented, replacing traditional single light source estimation with light distribution to support visual presentation in multi-light source scenarios.
Physics-Augmented Autoencoder for 3D Skeleton-Based Gait Recognition
Hongji Guo (Rensselaer Polytechnic Institute), Qiang Ji (Rensselaer Polytechnic Institute)
RecognitionPose EstimationRecurrent Neural NetworkGraph Neural NetworkAuto EncoderSequentialPhysics Related
🎯 What it does: A Physical Augmented Autoencoder (PAA) is proposed, which encodes three-dimensional skeleton sequences into generalized joint poses and forces, and reconstructs them through Lagrangian dynamics decoding;
Physics-Driven Turbulence Image Restoration with Stochastic Refinement
Ajay Jaiswal (University of Texas at Austin), Zhangyang Wang (Purdue University)
RestorationSuper ResolutionTransformerDiffusion modelImagePhysics Related
🎯 What it does: This paper proposes a method that directly integrates a physical layer atmospheric turbulence simulator into the training loop of a deep recovery network (PiRN), and adds a conditional diffusion model for post-processing (PiRN-SR) to enhance perceptual quality.
PIDRo: Parallel Isomeric Attention with Dynamic Routing for Text-Video Retrieval
Peiyan Guan (University of Hong Kong), Edmund Y. Lam (University of Hong Kong)
RetrievalTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: A text-video retrieval model based on CLIP is proposed and implemented with a parallel heterogeneous attention module (S-T frame branch and T-S patch branch) and a dynamic routing mechanism to enhance the video encoder and text word vectors, thereby achieving finer-grained cross-modal alignment and retrieval.
PIRNet: Privacy-Preserving Image Restoration Network via Wavelet Lifting
Xin Deng (Beihang University), Mai Xu (Beihang University)
RestorationSuper ResolutionSafty and PrivacyAuto EncoderImage
🎯 What it does: This paper proposes a reversible steganography and restoration network based on lifting transformation (PIRNet), achieving image denoising, deblurring, and super-resolution restoration tasks for cloud privacy protection.
PivotNet: Vectorized Pivot Learning for End-to-end HD Map Construction
Wenjie Ding (MEGVII Technology), Chi Zhang (MEGVII Technology)
Autonomous DrivingTransformerPoint Cloud
🎯 What it does: This paper proposes an end-to-end HD map construction framework called PivotNet based on pivot points, which directly predicts the key point sequences of each map element, avoiding the traditional post-processing of segmentation followed by vectorization.
Pix2Video: Video Editing using Image Diffusion
Duygu Ceylan (Adobe Research), Niloy J. Mitra (University College London)
GenerationData SynthesisDiffusion modelVideoText
🎯 What it does: Utilizing a pre-trained image diffusion model for text-driven editing of videos without the need for any additional training or fine-tuning;
Pixel Adaptive Deep Unfolding Transformer for Hyperspectral Image Reconstruction
Miaoyu Li (Beijing Institute of Technology), Yulun Zhang (ETH Zurich)
RestorationTransformerImage
🎯 What it does: This paper proposes a Pixel Adaptive Deep Unfolding Transformer (PADUT) for reconstructing 3D hyperspectral image (HSI) cubes from CASSI coded snapshot spectral images, improving the numerical updates, prior learning, and stage interaction of traditional unfolding frameworks.
Pixel-Aligned Recurrent Queries for Multi-View 3D Object Detection
Yiming Xie (Northeastern University), Julian Straub (Meta Reality Labs Research)
Object DetectionTransformerPoint Cloud
🎯 What it does: This paper proposes PARQ (Pixel-Aligned Recurrent Queries), a multi-view 3D object detection framework based on Transformer, which continuously updates the 3D positions using pixel-aligned query points during the recursive process and outputs 3D bounding boxes.
Pixel-Wise Contrastive Distillation
Junqiang Huang (Shopee), Zichao Guo (Shopee)
Object DetectionSegmentationKnowledge DistillationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes Pixel-Wise Contrastive Distillation (PCD), a self-supervised knowledge distillation framework aimed at dense prediction tasks, which transfers pixel features from the teacher model to the student model through pixel-level contrastive learning.
PlanarTrack: A Large-scale Challenging Benchmark for Planar Object Tracking
Xinran Liu (Institute of Software, Chinese Academy of Sciences), Heng Fan
Object TrackingConvolutional Neural NetworkVideoBenchmark
🎯 What it does: This paper constructs the first large-scale benchmark for planar object tracking in complex scenarios, called PlanarTrack, and evaluates ten planar trackers on this benchmark, while also introducing PlanarTrack BB for experiments on planar object localization with general trackers.
PlaneRecTR: Unified Query Learning for 3D Plane Recovery from a Single View
Jingjia Shi (National University of Defense Technology), Kai Xu (National University of Defense Technology)
SegmentationDepth EstimationTransformerPoint Cloud
🎯 What it does: Designed PlaneRecTR, which unifies the detection, parameter estimation, segmentation, and depth prediction of single-view plane recovery into a single Transformer query learning framework.
PlankAssembly: Robust 3D Reconstruction from Three Orthographic Views with Learnt Shape Programs
Wentao Hu (Sun Yat-Sen University), Zihan Zhou (Manycore Tech Inc.)
GenerationTransformerMeshBenchmark
🎯 What it does: This paper proposes a framework that utilizes a Transformer seq2seq model to automatically convert three-view orthographic line drawings into structured 3D CAD models, and implements the assembly of cabinet furniture through a domain-specific shape program (Plank Assembly DSL).
Plausible Uncertainties for Human Pose Regression
Lennart Bramlage (Continental AG), Cristóbal Curio (Reutlingen University)
Pose EstimationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This study investigates methods for simultaneously quantifying uncertainty in human pose estimation and conducts experimental evaluations.
Pluralistic Aging Diffusion Autoencoder
Peipei Li (Beijing University of Posts and Telecommunications), Zhaofeng He (Institute of Automation, Chinese Academy of Sciences)
GenerationData SynthesisDiffusion modelAuto EncoderImageMultimodality
🎯 What it does: A diversified facial aging diffusion autoencoder (PADA) based on CLIP is proposed, achieving the generation of multimodal and diverse aging results conditioned on text or reference images.
PNI : Industrial Anomaly Detection using Position and Neighborhood Information
Jaehyeok Bae (Seoul National University), Seyun Kim (Gauss Labs Inc.)
Anomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an industrial defect detection and localization algorithm called PNI, which utilizes position and neighborhood information, combines multilayer perceptrons with histogram estimation of normal feature distribution, and trains a pixel-level refinement network using synthetic defect images.
PODA: Prompt-driven Zero-shot Domain Adaptation
Mohammad Fahes (Inria), Raoul de Charette (Inria)
Object DetectionSegmentationDomain AdaptationPrompt EngineeringContrastive LearningImage
🎯 What it does: This paper proposes Prompt-driven Zero-shot Domain Adaptation (PØDA), which utilizes CLIP text prompts to adapt source domain models to unseen target domains.
PODIA-3D: Domain Adaptation of 3D Generative Model Across Large Domain Gap Using Pose-Preserved Text-to-Image Diffusion
Gwanghyun Kim (Seoul National University), Se Young Chun (Seoul National University)
GenerationDomain AdaptationDiffusion modelImage
🎯 What it does: Using a pose-preserving text-to-image diffusion model (PPD) and a specialized-to-general sampling strategy, we adapt existing 3D generators (such as EG3D) for cross-domain applications, addressing the pose-shape mismatch problem caused by large domain gaps.
Poincare ResNet
Max van Spengler (University of Amsterdam), Pascal Mettes (University of Amsterdam)
ClassificationAnomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: An end-to-end residual network called Poincaré ResNet is proposed, which is fully implemented in the Poincaré ball model to directly learn visual representations in hyperbolic space at the pixel level.
Point Contrastive Prediction with Semantic Clustering for Self-Supervised Learning on Point Cloud Videos
Xiaoxiao Sheng (Shanghai Jiao Tong University), Hehe Fan (Zhejiang University)
RecognitionSegmentationTransformerContrastive LearningVideoPoint Cloud
🎯 What it does: A unified point-level contrastive learning framework called PointCPSC is proposed, which combines semantic clustering and positive sample neighbor selection to achieve self-supervised pre-training for point cloud videos.
Point-Query Quadtree for Crowd Counting, Localization, and More
Chengxin Liu (Huazhong University of Science and Technology), Tongliang Liu (University of Sydney)
RecognitionObject DetectionTransformerImage
🎯 What it does: Treating crowd counting as a decomposable point query process, we propose the Point Query Transformer (PET) model, which employs a point query quadtree and advanced rectangular window attention to support multi-tasks such as counting, localization, partial annotation learning, and point annotation refinement.
Point-SLAM: Dense Neural Point Cloud-based SLAM
Erik Sandström (ETH Zurich), Martin R. Oswald (KU Leuven)
Simultaneous Localization and MappingPoint Cloud
🎯 What it does: A dense RGB-D SLAM framework based on neural point clouds, called Point-SLAM, is proposed.
Point-TTA: Test-Time Adaptation for Point Cloud Registration Using Multitask Meta-Auxiliary Learning
Ahmed Hatem (University of Manitoba), Yang Wang (Concordia University)
Domain AdaptationMeta LearningContrastive LearningPoint Cloud
🎯 What it does: Proposes a point cloud registration testing adaptive framework based on multi-task meta auxiliary learning (Point-TTA).
Point2Mask: Point-supervised Panoptic Segmentation via Optimal Transport
Wentong Li (Zhejiang University), Lei Zhang (HongKong Polytechnical University)
Object DetectionSegmentationTransformerImage
🎯 What it does: A single-point supervised panoptic segmentation method called Point2Mask is proposed, which generates pseudo-masks through optimal transport and trains a panoptic segmentation network.
PointCLIP V2: Prompting CLIP and GPT for Powerful 3D Open-world Learning
Xiangyang Zhu (City University of Hong Kong), Peng Gao (Peking University)
ClassificationObject DetectionSegmentationTransformerLarge Language ModelPrompt EngineeringVision Language ModelPoint Cloud
🎯 What it does: This paper presents PointCLIP V2, which unifies the prompting methods of CLIP and GPT-3 to transfer the pre-trained vision-language model to the 3D point cloud domain, achieving unsupervised learning for zero/few-shot 3D classification, segmentation, and detection tasks.
PointDC: Unsupervised Semantic Segmentation of 3D Point Clouds via Cross-Modal Distillation and Super-Voxel Clustering
Zisheng Chen (South China University of Technology), Wenxiong kang
SegmentationKnowledge DistillationContrastive LearningPoint Cloud
🎯 What it does: The PointDC framework is proposed, achieving completely unsupervised 3D point cloud semantic segmentation.
PointMBF: A Multi-scale Bidirectional Fusion Network for Unsupervised RGB-D Point Cloud Registration
Mingzhi Yuan (Fudan University), Manning Wang (Fudan University)
Domain AdaptationConvolutional Neural NetworkImagePoint Cloud
🎯 What it does: This paper proposes a multi-scale bidirectional fusion network called PointMBF for unsupervised RGB-D point cloud registration, which can simultaneously utilize the complementary information of RGB images and depth point clouds.
PointOdyssey: A Large-Scale Synthetic Dataset for Long-Term Point Tracking
Yang Zheng (Stanford University), Leonidas J. Guibas (Stanford University)
Object TrackingData SynthesisConvolutional Neural NetworkVideo
🎯 What it does: A large synthetic dataset called PointOdyssey and an improved point tracking method PIPs++ are proposed, focusing on fine-grained point tracking over long durations;
PolicyCleanse: Backdoor Detection and Mitigation for Competitive Reinforcement Learning
Junfeng Guo (University of Maryland), Cong Liu (University of California Riverside)
Reinforcement LearningSequential
🎯 What it does: Detecting and eliminating backdoor attacks in competitive reinforcement learning
Ponder: Point Cloud Pre-training via Neural Rendering
Di Huang (University of Sydney), Wanli Ouyang (University of Sydney)
Object DetectionSegmentationRepresentation LearningNeural Radiance FieldContrastive LearningImagePoint Cloud
🎯 What it does: By pre-training RGB-D images through differentiable neural rendering, the point cloud encoder learns to generate feature vectors that contain both geometric and appearance information.
Pose-Free Neural Radiance Fields via Implicit Pose Regularization
Jiahui Zhang (Nanyang Technological University), Shijian Lu (Nanyang Technological University)
GenerationPose EstimationNeural Radiance FieldGenerative Adversarial NetworkImage
🎯 What it does: A Pose-free NeRF (IR-NeRF) is proposed, which trains NeRF on multi-view images without poses by constructing a scene codebook and utilizing implicit pose regularization.
PoseDiffusion: Solving Pose Estimation via Diffusion-aided Bundle Adjustment
Jianyuan Wang (University of Oxford), David Novotny (Meta AI)
Pose EstimationOptimizationTransformerDiffusion modelImageVideo
🎯 What it does: A Bundle Adjustment method called PoseDiffusion based on diffusion models is proposed for estimating camera intrinsic and extrinsic parameters in multi-view images. This method gradually refines random camera poses through reverse sampling during the diffusion process, guided by geometric constraints.
PoseFix: Correcting 3D Human Poses with Natural Language
Ginger Delmas (Institut de Robotica i Informatica Industrial), Grégory Rogez (NAVER LABS Europe)
GenerationPose EstimationTransformerLarge Language ModelAuto EncoderText
🎯 What it does: This study proposes the PoseFix dataset, establishing a correspondence between 3D human pose pairs and natural language correction descriptions, and based on this, introduces two tasks: text-driven pose editing and pose difference text generation.
PourIt!: Weakly-Supervised Liquid Perception from a Single Image for Visual Closed-Loop Robotic Pouring
Haitao Lin (Fudan University), Xiangyang Xue (Fudan University)
SegmentationPose EstimationRobotic IntelligenceTransformerContrastive LearningImagePoint Cloud
🎯 What it does: A visual closed-loop robotic pouring framework named PourIt! is proposed, which achieves weakly supervised liquid segmentation using image-level labels and CAM+ feature comparison, and reconstructs the 3D point cloud of the liquid by estimating the pose of the source container, providing visual feedback on the distance from the liquid to the target container.
PPR: Physically Plausible Reconstruction from Monocular Videos
Gengshan Yang (Carnegie Mellon University), Deva Ramanan (Carnegie Mellon University)
Neural Radiance FieldVideo
🎯 What it does: Combining differentiable rendering and differentiable physical simulation, we achieve physically plausible dynamic 3D model reconstruction from monocular video.
Practical Membership Inference Attacks Against Large-Scale Multi-Modal Models: A Pilot Study
Myeongseob Ko (Virginia Tech), Ruoxi Jia (Virginia Tech)
Adversarial AttackTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes membership inference attacks against large multimodal models, including baseline attacks based on cosine similarity, enhanced attacks, and weakly supervised attacks.
PRANC: Pseudo RAndom Networks for Compacting Deep Models
Parsa Nooralinejad (University of California), Hamed Pirsiavash (University of California)
ClassificationCompressionConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: Reparameterizing deep models as a linear combination of several randomly initialized, frozen base networks only requires saving the random seed and combination coefficients to restore the model.
Pre-training Vision Transformers with Very Limited Synthesized Images
Ryo Nakamura (National Institute of Advanced Industrial Science and Technology), Nakamasa Inoue (National Institute of Advanced Industrial Science and Technology)
ClassificationObject DetectionSegmentationData SynthesisTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: A scheme was developed to pre-train a visual Transformer with minimal synthetic images by generating a fractal database (OFDB) that contains only one image per category and employing data augmentation during the pre-training phase.
Pre-Training-Free Image Manipulation Localization through Non-Mutually Exclusive Contrastive Learning
Jizhe Zhou (Sichuan University), Wentao Feng (Sichuan University)
Anomaly DetectionConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: A framework for image forgery localization based on Non-Exclusive Contrastive Learning (NCL) is proposed, which does not rely on pre-trained data and directly trains deep networks from the raw dataset.
Predict to Detect: Prediction-guided 3D Object Detection using Sequential Images
Sanmin Kim (KAIST), Dongsuk Kum (KAIST)
Object DetectionAutonomous DrivingTransformerPoint CloudSequential
🎯 What it does: This paper proposes an end-to-end multi-frame 3D object detection framework called P2D, which combines prediction and detection. It utilizes past frames to predict the object information of the current frame and performs feature aggregation to improve detection accuracy.
Preface: A Data-driven Volumetric Prior for Few-shot Ultra High-resolution Face Synthesis
Marcel C. Bühler (ETH Zurich), Abhimitra Meka (Google)
GenerationData SynthesisNeural Radiance FieldImage
🎯 What it does: A NeRF prior model based on identity conditions is proposed, capable of quickly reconstructing and rendering ultra-high-resolution new views of faces at 4K level from just two or three images.
Preparing the Future for Continual Semantic Segmentation
Zihan Lin (University of Science and Technology of China), Yixin Zhang (University of Science and Technology of China)
SegmentationKnowledge DistillationContrastive LearningImage
🎯 What it does: This paper proposes a method for pre-learning knowledge of future unknown categories in continuous semantic segmentation to address the challenges of learning new classes and the problem of catastrophic forgetting.
Preserve Your Own Correlation: A Noise Prior for Video Diffusion Models
Songwei Ge (NVIDIA), Yogesh Balaji (NVIDIA)
GenerationData SynthesisDiffusion modelVideoText
🎯 What it does: Based on a pre-trained text-to-image diffusion model, a video diffusion model was designed and finetuned to generate high-quality, temporally consistent videos from text.
Preserving Modality Structure Improves Multi-Modal Learning
Sirnam Swetha (University of Central Florida), Mubarak Shah (University of Central Florida)
RetrievalDomain AdaptationTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: A method for maintaining semantic structure consistency (SSPC) is proposed, which utilizes multiple anchor points to preserve modality-specific semantic relationships in multimodal pre-training, thereby enhancing cross-domain generalization capabilities.
Preserving Tumor Volumes for Unsupervised Medical Image Registration
Qihua Dong (City University of Hong Kong), Jing Liao (City University of Hong Kong)
SegmentationOptimizationConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A two-stage unsupervised medical image registration method is proposed, which preserves tumor volume while maintaining image similarity.
PreSTU: Pre-Training for Scene-Text Understanding
Jihyung Kil (Ohio State University), Radu Soricut (Google Research)
RecognitionTransformerVision Language ModelImageText
🎯 What it does: A new pre-training method called PRESTU is proposed to enhance the performance of visual-language models in scene text understanding tasks.
Pretrained Language Models as Visual Planners for Human Assistance
Dhruvesh Patel (Meta), Ruta Desai (Meta)
SegmentationGenerationTransformerLarge Language ModelVision Language ModelVideoMultimodality
🎯 What it does: This paper proposes the Visual Planning for Assistance (VPA) task and constructs the VLaMP model based on a two-stage approach of video segmentation and prediction to generate the next action sequence under the conditions of user-defined goals and video progress.
Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models
Zangwei Zheng (National University of Singapore), Yang You (National University of Singapore)
Knowledge DistillationRepresentation LearningTransformerVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: In continual learning, a method called ZSCL is proposed to protect the zero-shot transfer capability of pre-trained vision-language models (such as CLIP) while accommodating the acquisition of new knowledge when learning new tasks.
Prior-guided Source-free Domain Adaptation for Human Pose Estimation
Dripta S. Raychaudhuri (University of California), Amit K. Roy-Chowdhury (University of California)
Pose EstimationDomain AdaptationContrastive LearningImageVideo
🎯 What it does: This paper proposes an unsupervised domain adaptation method for 2D human pose estimation (POST) that does not require source data, achieving transfer from the source domain to the target domain through self-supervised pseudo-label learning.
PRIOR: Prototype Representation Joint Learning from Medical Images and Reports
Pujin Cheng (Southern University of Science and Technology), Xiaoying Tang (Southern University of Science and Technology)
ClassificationObject DetectionSegmentationRetrievalRepresentation LearningTransformerContrastive LearningImageTextMultimodalityBiomedical Data
🎯 What it does: A Prototype Representation Joint Learning (PRIOR) framework is proposed for paired medical imaging and reporting data, combining global and local alignment, a sentence-level prototype memory bank, and cross-modal conditional reconstruction to achieve finer-grained cross-modal representation learning.
Priority-Centric Human Motion Generation in Discrete Latent Space
Hanyang Kong (National University of Singapore), Xinchao Wang (National University of Singapore)
GenerationTransformerReinforcement LearningDiffusion modelTextMultimodality
🎯 What it does: A priority-centered discrete diffusion model (M2DM) has been constructed to generate high-quality and diverse 3D human actions based on natural language descriptions.
Privacy Preserving Localization via Coordinate Permutations
Linfei Pan (ETH Zurich), Marc Pollefeys (ETH Zurich)
Pose EstimationSafty and PrivacyImage
🎯 What it does: This study proposes a coordinate displacement scheme to recover the original point positions while ensuring the privacy of query images or maps, thus achieving precise camera positioning under privacy protection.
Privacy-Preserving Face Recognition Using Random Frequency Components
Yuxi Mi (Fudan University), Shuigeng Zhou (Fudan University)
RecognitionSafty and PrivacyConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper studies a privacy-preserving facial recognition method called PartialFace, which is achieved by randomly selecting high-frequency components in the frequency domain.
Probabilistic Human Mesh Recovery in 3D Scenes from Egocentric Views
Siwei Zhang (ETH Zurich), Siyu Tang (ETH Zurich)
GenerationPose EstimationGraph Neural NetworkDiffusion modelMesh
🎯 What it does: EgoHMR is proposed, a scene-conditioned diffusion model for recovering 3D human meshes from single-frame first-person images, specifically addressing the issue of body truncation.
Probabilistic Modeling of Inter- and Intra-observer Variability in Medical Image Segmentation
Arne Schmidt (Universidad de Granada), Rafael Molina (Universidad de Granada)
SegmentationConvolutional Neural NetworkAuto EncoderImageMagnetic Resonance Imaging
🎯 What it does: A probabilistic deep learning model named Pionono is proposed for modeling inter-observer and intra-observer variability in medical image segmentation, trained end-to-end through variational inference.
Probabilistic Precision and Recall Towards Reliable Evaluation of Generative Models
Dogyun Park (Korea University), Suhyun Kim (Korea Institute of Science and Technology)
GenerationData SynthesisDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: A probability-based precision and recall metric, P-precision and P-recall, is proposed to more reliably evaluate the fidelity and diversity of generative models.
Probabilistic Triangulation for Uncalibrated Multi-View 3D Human Pose Estimation
Boyuan Jiang (Institute of Computing Technology, Chinese Academy of Sciences), Shihong Xia (Institute of Computing Technology, Chinese Academy of Sciences)
Pose EstimationImageVideo
🎯 What it does: This paper proposes a Probabilistic Triangulation module that can achieve 3D human pose estimation in uncalibrated multi-view scenarios.
ProbVLM: Probabilistic Adapter for Frozen Vison-Language Models
Uddeshya Upadhyay (University of Tübingen), Zeynep Akata (MPI for Intelligent Systems)
RetrievalTransformerVision Language ModelDiffusion modelImageText
🎯 What it does: A post-processing adapter called ProbVLM is designed to convert the deterministic embeddings of frozen visual-language models (such as CLIP and BLIP) into probability distributions, thereby obtaining uncertainty estimates without retraining the models.
Progressive Spatio-Temporal Prototype Matching for Text-Video Retrieval
Pandeng Li (University of Science and Technology of China), Yongdong Zhang (DAMO Academy Alibaba Group)
RetrievalTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes a text-video retrieval framework called ProST based on advanced spatial-temporal prototype matching. It first generates spatial prototypes to match local objects and phrases, and then generates temporal prototypes to match events and sentences, achieving multi-granularity alignment.
Prompt Switch: Efficient CLIP Adaptation for Text-Video Retrieval
Chaorui Deng (Australia Institute of Machine Learning, University of Adelaide), Qi Wu (Australia Institute of Machine Learning, University of Adelaide)
RetrievalTransformerPrompt EngineeringContrastive LearningVideoTextMultimodality
🎯 What it does: In text-video retrieval, global semantic modeling of videos is achieved using the image encoder of CLIP through a switchable three-dimensional Prompt Cube, and fine-grained semantics are enhanced with auxiliary video subtitle objectives.
Prompt Tuning Inversion for Text-driven Image Editing Using Diffusion Models
Wenkai Dong (Baidu), Shumin Han (Baidu)
Image TranslationGenerationPrompt EngineeringDiffusion modelImageText
🎯 What it does: This paper proposes a text-driven image editing method based on diffusion models, utilizing Prompt Tuning Inversion to encode the original image information into a learnable conditional embedding, which is then linearly interpolated with the target text embedding to achieve image editing.
Prompt-aligned Gradient for Prompt Tuning
Beier Zhu (Nanyang Technological University), Hanwang Zhang (Nanyang Technological University)
ClassificationDomain AdaptationTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: ProGrad is proposed, a prompt tuning method for CLIP that utilizes a gradient alignment mechanism to avoid overfitting and knowledge forgetting in low-sample scenarios.
PromptCap: Prompt-Guided Image Captioning for VQA with GPT-3
Yushi Hu (University of Washington), Jiebo Luo (University of Rochester)
GenerationRetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes PROMPTCAP, a visual description model controlled by natural language prompts, which transforms images into customized textual descriptions tailored to questions, enabling black-box large language models like GPT-3 to understand images and perform knowledge-driven visual question answering.
PromptStyler: Prompt-driven Style Generation for Source-free Domain Generalization
Junhyeong Cho (ADD), Suha Kwak (POSTECH)
GenerationDomain AdaptationTransformerPrompt EngineeringVision Language ModelContrastive LearningImageText
🎯 What it does: This paper proposes a passive domain generalization method called PromptStyler, which learns variable style word vectors in the joint visual-language space of CLIP to synthesize features of various styles using only text prompts, thereby achieving source-free domain generalization without using any image data.
ProPainter: Improving Propagation and Transformer for Video Inpainting
Shangchen Zhou (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
RestorationTransformerOptical FlowVideo
🎯 What it does: ProPainter framework is proposed, which combines dual-domain propagation (image domain + feature domain) with mask-guided sparse Transformer to address issues such as long-distance correspondence, texture consistency, and inefficiency in video inpainting.
ProtoFL: Unsupervised Federated Learning via Prototypical Distillation
Hansol Kim (KakaoBank Corp), Changick Kim (KAIST)
Federated LearningKnowledge DistillationRepresentation LearningConvolutional Neural NetworkFlow-based ModelContrastive LearningImage
🎯 What it does: Proposes the ProtoFL two-stage unsupervised federated learning framework, which first distills global features through the prototype representation of an offline pre-trained model, and then uses normalized flow to train a one-class classifier on each client to address one-class classification tasks in extremely non-i.i.d. federated environments.
ProtoTransfer: Cross-Modal Prototype Transfer for Point Cloud Segmentation
Pin Tang (Shanghai Jiao Tong University), Chao Ma (Shanghai Jiao Tong University)
SegmentationAutonomous DrivingKnowledge DistillationConvolutional Neural NetworkImagePoint Cloud
🎯 What it does: This paper proposes a cross-modal prototype transfer method (ProtoTransfer), which constructs a category-level prototype library to transfer the fused LiDAR and image feature knowledge to all point cloud features, achieving high performance in single-modal (LiDAR only) point cloud semantic segmentation.
Prototype Reminiscence and Augmented Asymmetric Knowledge Aggregation for Non-Exemplar Class-Incremental Learning
Wuxuan Shi (Wuhan University), Mang Ye (Wuhan University)
ClassificationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes two mechanisms: prototype recall and enhanced asymmetric knowledge aggregation, aimed at the scenario of class-incremental learning without samples, which preserves old class knowledge while improving the learning effectiveness of new classes.
Prototype-based Dataset Comparison
Nanne van Noord (University of Amsterdam)
ClassificationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: The paper proposes the ProtoSim module, which learns integrated prototypes in Vision Transformers, achieving prototype learning and dataset comparison across different datasets.
Prototypes-oriented Transductive Few-shot Learning with Conditional Transport
Long Tian (Xidian University), Bo Chen (Xidian University)
ClassificationContrastive LearningImage
🎯 What it does: This study investigates transductive few-shot learning with class imbalance, proposing the PUTM model that achieves unbiased statistical transfer through Conditional Transport (CT), improving prototype generation and classification.
Prototypical Kernel Learning and Open-set Foreground Perception for Generalized Few-shot Semantic Segmentation
Kai Huang (Alibaba Group), Yutao Gao (Alibaba Group)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: A general few-shot semantic segmentation framework is proposed to simultaneously address the separation of base class and new class representations and the embedding bias problem, with the core being Prototype Kernel Learning (PKL) and Open Set Foreground Perception (FCP) modules, and the predictions of both are fused through Conditional Bias-Based Inference (CBBI).
Prototypical Mixing and Retrieval-Based Refinement for Label Noise-Resistant Image Retrieval
Xinlong Yang (Peking University), Xiao Luo (University of California, Los Angeles)
RetrievalConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper addresses the problem of image retrieval with label noise and proposes a new method called TITAN, which can simultaneously correct label errors and suppress the model's overfitting to noisy samples, thereby improving retrieval performance.
Proxy Anchor-based Unsupervised Learning for Continuous Generalized Category Discovery
Hyungmin Kim (Korea Advanced Institute of Science and Technology), Junmo Kim (Korea Advanced Institute of Science and Technology)
ClassificationKnowledge DistillationRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Unsupervised category incremental learning is performed on an unlabeled joint dataset, automatically distinguishing between known and new categories, discovering new categories during continuous incremental phases, while suppressing catastrophic forgetting.
Prune Spatio-temporal Tokens by Semantic-aware Temporal Accumulation
Shuangrui Ding (Shanghai Jiao Tong University), Qi Tian (Huawei Cloud)
RecognitionOptimizationComputational EfficiencyTransformerVideo
🎯 What it does: This paper proposes a Semantic-Aware Temporal Accumulation (STA) score for pruning spatiotemporal tokens in video Transformers.
Pseudo Flow Consistency for Self-Supervised 6D Object Pose Estimation
Yang Hai (Xidian University), Yinlin Hu (MagicLeap)
Pose EstimationOptical FlowImage
🎯 What it does: Proposes a self-supervised 6D object pose estimation framework that only uses RGB images;
Pseudo-label Alignment for Semi-supervised Instance Segmentation
Jie Hu (Xiamen University), Rongrong Ji (Xiamen University)
Object DetectionSegmentationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes a semi-supervised instance segmentation framework called PAIS, which utilizes pseudo-labels to train on unlabeled images and improves the utilization of pseudo-labels through Dynamic Alignment Loss (DALoss).
PVT++: A Simple End-to-End Latency-Aware Visual Tracking Framework
Bowen Li (Carnegie Mellon University), Changhong Fu (National University of Singapore)
Object TrackingVideoBenchmark
🎯 What it does: An end-to-end predictive visual tracking framework PVT++ is proposed, which compensates for online inference delays by adding a lightweight prediction module after the tracker, allowing the tracking results to correspond to the real-world state when the camera finishes processing.
Pyramid Dual Domain Injection Network for Pan-sharpening
Xuanhua He (University of Science and Technology of China), Man Zhou (Nanyang Technological University)
RestorationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A dual-domain pyramid injection network is designed and implemented, utilizing multi-scale information from PAN images in both spatial and frequency domains to inject into LRMS images, thereby achieving high-quality pan-sharpening.
Q-Diffusion: Quantizing Diffusion Models
Xiuyu Li (University of California Berkeley), Kurt Keutzer (University of California Berkeley)
GenerationCompressionConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Q-Diffusion is proposed, a post-training quantization (PTQ) method specifically designed for diffusion models, capable of compressing the noise estimation network to 4-bit weight quantization while maintaining generation quality.
QD-BEV : Quantization-aware View-guided Distillation for Multi-view 3D Object Detection
Yifan Zhang (Nanjing University), Shanghang Zhang (Peking University)
Object DetectionCompressionAutonomous DrivingKnowledge DistillationTransformerSupervised Fine-TuningImagePoint Cloud
🎯 What it does: This paper proposes a low-bit-width quantization method for multi-view 3D object detection called QD-BEV, which enhances the stability of quantization training and performance recovery through View-Guided Distillation (VGD), ultimately achieving a significantly compressed BEV model on the nuScenes dataset.
Quality Diversity for Visual Pre-Training
Ruchika Chavhan (University of Edinburgh), Timothy Hospedales (University of Edinburgh)
ClassificationObject DetectionSegmentationPose EstimationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A quality diversity pre-training framework QD4V is proposed, which trains a set of features that are both high-quality and diversified in terms of different data augmentations (in)variance, achieving better performance in downstream tasks through stacked fusion.