ECCV 2024 Papers — Page 16
European Conference on Computer Vision · 2387 papers
OmniSat: Self-Supervised Modality Fusion for Earth Observation
Guillaume Astruc (Univ Gustave Eiffel), Loic Landrieu (Univ Gustave Eiffel)
ClassificationRepresentation LearningConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningMultimodalityTime SeriesBenchmarkAgriculture Related
🎯 What it does: Propose OmniSat, a self-supervised multimodal fusion framework designed to integrate representations from different Earth observation sensors (VHR aerial imagery, optical time series, radar time series), and enhance downstream task performance through self-supervised pre-training in multimodal settings.
OmniSSR: Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model
Runyi Li (Peking University), Jian Zhang (Peking University)
Super ResolutionDiffusion modelImage
🎯 What it does: This paper proposes a zero-shot panoramic image super-resolution method called OmniSSR, which utilizes Stable Diffusion for super-resolution and enhances image quality through iterative ERP and TP projection interaction with gradient decomposition correction.
Omniview-Tuning: Boosting Viewpoint Invariance of Vision-Language Pre-training Models
Shouwei Ruan (Beihang University), Xingxing Wei (Beihang University)
TransformerSupervised Fine-TuningVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose a new framework, Omniview-Tuning (OVT), which significantly enhances the model's robustness under three-dimensional viewpoint shifts by constructing the MVCap dataset with over four million multi-view图文 pairs and performing parameter-efficient viewpoint-invariant fine-tuning on the VLP model using this dataset.
OMR: Occlusion-Aware Memory-Based Refinement for Video Lane Detection
Dongkwon Jin, Chang-Su Kim
Object DetectionSegmentationAutonomous DrivingConvolutional Neural NetworkRecurrent Neural NetworkVideo
🎯 What it does: Proposed a real-time lane detection method based on a multi-task deep convolutional network
On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition
Zihu Wang (University of California Santa Barbara), Peng Li (Carnegie Mellon University)
RecognitionData SynthesisRepresentation LearningAuto EncoderContrastive LearningImage
🎯 What it does: Propose a self-supervised learning method that generates synthetic data pairs in the latent space, utilizing a decoder to reconstruct original and perturbed features, emphasizing discriminative features in fine-grained visual recognition.
On Pretraining Data Diversity for Self-Supervised Learning
Hasan Abed Al Kader Hammoud (KAUST University of Oxford), Bernard Ghanem
ClassificationRecognitionRepresentation LearningData-Centric LearningContrastive LearningImage
🎯 What it does: Under a fixed computational budget, the study systematically evaluates and compares the impact of pre-training data with different scales and distributions on the performance of self-supervised learning (SSL).
On Spectral Properties of Gradient-based Explanation Methods
Amir Mehrpanah (Royal Institute of Technology), Hossein Azizpour (Royal Institute of Technology)
Explainability and InterpretabilityImage
🎯 What it does: This paper formally analyzes gradient-based explanation methods from a probabilistic and spectral perspective, revealing their spectral bias and sensitivity to perturbation hyperparameters, and proposes two improvement schemes (optimal perturbation scale and SpectralLens aggregation method).
On the Approximation Risk of Few-Shot Class-Incremental Learning
Xuan Wang (Tianjin University), Jungong Han (University of Sheffield)
ClassificationMeta LearningTransformerImageBenchmark
🎯 What it does: This paper conducts an in-depth analysis of the approximate risk of few-shot incremental learning (FSCIL) from the perspective of statistical learning theory, and proposes a series of operable training and design guidelines based on this analysis;
On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy
Letian Huang (Nanjing University), Yanwen Guo (Nanjing University)
OptimizationGaussian SplattingPoint Cloud
🎯 What it does: This paper conducts a theoretical analysis of projection errors in 3D Gaussian Splatting and proposes an optimal projection strategy to enhance rendering quality.
On the Evaluation Consistency of Attribution-based Explanations
Jiarui Duan (Zhejiang University), Jie Song (Zhejiang University)
Explainability and InterpretabilityConvolutional Neural NetworkImageBenchmark
🎯 What it does: Built the Meta-Rank platform for systematic evaluation of attribution methods in the image domain;
On the Topology Awareness and Generalization Performance of Graph Neural Networks
Junwei Su (University of Hong Kong), Chuan Wu (University of Hong Kong)
ClassificationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Propose a topology-aware framework based on approximate metric embedding, investigate the generalization performance of GNNs on different structural subgroups, and validate the framework using shortest path distance.
On the Utility of 3D Hand Poses for Action Recognition
Md Salman Shamil (National University of Singapore), Angela Yao (National University of Singapore)
RecognitionTransformerVideoMultimodality
🎯 What it does: Propose HandFormer, a lightweight multimodal Transformer that combines dense 3D hand pose with sparse RGB for hand-object interaction action recognition.
On the Viability of Monocular Depth Pre-training for Semantic Segmentation
Dong Lao (University of California Los Angeles), Stefano Soatto (University of California Los Angeles)
SegmentationDepth EstimationAutonomous DrivingConvolutional Neural NetworkSupervised Fine-TuningAuto EncoderOptical FlowImageVideoPoint Cloud
🎯 What it does: Can monocular depth estimation be used as unsupervised pre-training, followed by transferring the model to semantic segmentation tasks, while systematically evaluating the impact of different pre-training methods, data scale, network architecture, and other factors on segmentation performance?
On the Vulnerability of Skip Connections to Model Inversion Attacks
Jun Hao Koh (Singapore University of Technology and Design), Ngai-Man Cheung (Singapore University of Technology and Design)
Safty and PrivacyAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Investigated the vulnerability of skip connections to model inversion attacks (MI), and proposed a novel anti-MI network architecture based on removing the final stage skip connections (RoLSS).
On-the-fly Category Discovery for LiDAR Semantic Segmentation
Hyeonseong Kim (KAIST), Kuk-Jin Yoon (KAIST)
SegmentationAutonomous DrivingConvolutional Neural NetworkContrastive LearningPoint Cloud
🎯 What it does: This paper proposes a framework for on-the-fly category discovery (OCDSS) in LiDAR semantic segmentation, which can real-time identify and segment both known and unknown categories during testing, using only known categories for training.
One-Shot Diffusion Mimicker for Handwritten Text Generation
Gang Dai (South China University Of Technology), Shuangping Huang (South China University Of Technology)
GenerationDiffusion modelContrastive LearningImage
🎯 What it does: Propose a One-Shot Diffusion Mimicker (One-DM) model that can mimic the handwriting style of any writer and generate corresponding handwritten text images using only a single sample.
One-stage Prompt-based Continual Learning
Youngeun Kim (Yale University), Priyadarshini Panda (Yale University)
Computational EfficiencyRepresentation LearningMeta LearningTransformerPrompt EngineeringImage
🎯 What it does: Proposed a single-stage prompt-based continual learning (OS-Prompt) framework that directly uses CLS embeddings from Transformer intermediate layers as queries, eliminating the traditional two-stage ViT query process;
OneRestore: A Universal Restoration Framework for Composite Degradation
Yu Guo (Wuhan University of Technology), Shengfeng He (Singapore Management University)
RestorationTransformerContrastive LearningImage
🎯 What it does: Propose a general image restoration framework based on Transformer called OneRestore, which can simultaneously handle multiple composite degradations such as low light, fog, rain, and snow.
OneTrack: Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers
Qitai Wang (University of Chinese Academy of Sciences), Zhaoxiang Zhang (University of Chinese Academy of Sciences)
Object DetectionObject TrackingTransformerImage
🎯 What it does: Proposes OneTrack, a model that achieves end-to-end 3D detection and tracking under a unified object feature representation.
OneVOS: Unifying Video Object Segmentation with All-in-One Transformer Framework
Wanyun Li (Fudan University), Wenqiang Zhang (Fudan University)
SegmentationConvolutional Neural NetworkTransformerVideoBenchmark
🎯 What it does: Propose OneVOS, a full-process video object segmentation framework based on Vision Transformer, which unifies traditional phased modules such as feature extraction, matching, memory management, and multi-object aggregation into a single end-to-end trainable model.
Online Continuous Generalized Category Discovery
Keon-Hee Park (Kyung Hee University), Gyeong-Moon Park (Kyung Hee University)
ClassificationAnomaly DetectionTransformerScore-based ModelContrastive LearningImage
🎯 What it does: This paper proposes the Online Continuous General Category Discovery (OCGCD) scenario and introduces the DEAN method, which achieves online unknown category detection, pseudo-label generation, and continual learning through energy-guided discovery and variance feature enhancement.
Online Temporal Action Localization with Memory-Augmented Transformer
Youngkil Song (Pohang University of Science and Technology), Suha Kwak (Pohang University of Science and Technology)
Object DetectionTransformerVideo
🎯 What it does: Proposes an online temporal action localization model called Memory-Augmented Transformer (MATR), which captures long-term context in streaming video using a memory queue and achieves precise action instance localization and classification through an end-to-end Transformer structure.
Online Vectorized HD Map Construction using Geometry
Zhixin Zhang (Beijing Institute of Technology), Xiangyu Yue (Chinese University of Hong Kong)
Autonomous DrivingTransformerImage
🎯 What it does: Proposes an end-to-end online vectorization framework for high-definition map construction called GeMap, which learns the Euclidean shape and relationships of map instances through geometric representation.
Online Video Quality Enhancement with Spatial-Temporal Look-up Tables
Zefan Qu (Tongji University), Cairong Zhao (Microsoft Research Asia)
RestorationCompressionConvolutional Neural NetworkVideo
🎯 What it does: Propose a real-time video quality enhancement framework named STLVQE, which achieves artifact removal and quality improvement for online compressed videos through module-agnostic feature extraction, time caching, lightweight offset prediction, and spatial-temporal lookup tables (ST-LUT).
Online Zero-Shot Classification with CLIP
Qi Qian (Alibaba Group), Juhua Hu (University of Washington)
ClassificationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: Propose an online zero-shot classification method called OnZeta, which can classify continuously arriving images in real-time and dynamically update the model without storing samples;
OP-Align: Object-level and Part-level Alignment for Self-supervised Category-level Articulated Object Pose Estimation
Yuchen Che (Tokyo Institute of Technology), Asako Kanezaki (Tokyo Institute of Technology)
Pose EstimationConvolutional Neural NetworkPoint Cloud
🎯 What it does: Proposed a self-supervised OP-Align method that achieves category-level articulated object pose estimation using single-frame point clouds, and realizes object-level and part-level alignment with canonical reconstruction.
Open Panoramic Segmentation
Junwei Zheng (Karlsruhe Institute of Technology), Rainer Stiefelhagen (Karlsruhe Institute of Technology)
SegmentationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: Proposed the Open Panoramic Segmentation (OPS) task and designed the OOOPS model to achieve open-vocabulary panoramic segmentation without full panoramic annotations or zero-shot panoramic segmentation; the model is trained in an open-vocabulary manner on narrow field-of-view (pinhole) images and evaluated on wide field-of-view (360°) panoramic images.
Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation
Pengfei Wang (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
SegmentationKnowledge DistillationConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningPoint Cloud
🎯 What it does: Propose Geometry Guided Self-Distillation (GGSD), which leverages geometric priors to guide knowledge distillation from 2D pre-trained models and performs self-distillation within 3D networks to achieve open-vocabulary 3D scene understanding.
Open Vocabulary Multi-Label Video Classification
Rohit Gupta (University of Central Florida), Trishul A Chilimbi
ClassificationTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoText
🎯 What it does: Achieve open-vocabulary multi-label classification for various concepts (entities, actions, scenes, etc.) in videos by proposing an end-to-end training framework that combines large language models (LLMs) with pre-trained CLIP vision-language models.
Open-Set Biometrics: Beyond Good Closed-Set Models
Yiyang Su (Michigan State University), Xiaoming Liu (Michigan State University)
RecognitionContrastive LearningImageVideoMultimodality
🎯 What it does: A loss function that simultaneously optimizes recognition and detection, as well as minimizes the relative threshold, is proposed to enhance the FNIR@FPIR performance for open-set biometric recognition.
Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning
Dexuan Zhang (The University of Tokyo), Tatsuya Harada (The University of Tokyo)
Domain AdaptationImage
🎯 What it does: This paper proposes an end-to-end open-domain adaptation method that combines positive and unlabeled (PU) learning theory with joint error theory to derive a tight upper bound on the target risk directly.
Open-Set Recognition in the Age of Vision-Language Models
Dimity Miller (Queensland University of Technology), Keita Mason (Queensland University of Technology)
RecognitionObject DetectionTransformerVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Systematically evaluated the vulnerability of vision-language models (VLMs) in open-set recognition, and proposed updated definitions, benchmarks, and evaluation protocols for open-set problems specific to VLMs.
Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models
Xiaoyu Zhu (Carnegie Mellon University), Andrew Gallagher (HKUST)
SegmentationKnowledge DistillationConvolutional Neural NetworkTransformerDiffusion modelPoint Cloud
🎯 What it does: Achieve open-vocabulary 3D semantic segmentation and visual localization without 3D labels by leveraging pre-trained text-image diffusion models.
Open-Vocabulary Camouflaged Object Segmentation
Youwei Pang (Dalian University of Technology), Huchuan Lu (Dalian University of Technology)
SegmentationTransformerPrompt EngineeringVision Language ModelImageText
🎯 What it does: Proposed the open-vocabulary camouflage object segmentation task, constructed a large-scale camouflage object dataset named OVCamo, and introduced a single-stage Transformer baseline model called OVCoser based on CLIP.
Open-Vocabulary RGB-Thermal Semantic Segmentation
GuoQiang Zhao, Tao Peng (Wuhan Textile University)
SegmentationTransformerPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Proposes OpenRSS, an end-to-end method capable of performing open-vocabulary semantic segmentation on RGB and thermal images.
Open-Vocabulary SAM: Segment and Recognize Twenty-thousand Classes Interactively
Haobo Yuan (S-Lab, Nanyang Technological University), Chen Change Loy (S-Lab, Nanyang Technological University)
RecognitionSegmentationKnowledge DistillationTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: This paper proposes Open-Vocabulary SAM, a unified encoder-decoder framework that combines CLIP and SAM to achieve interactive segmentation and recognition.
Open-World Dynamic Prompt and Continual Visual Representation Learning
Youngeun Kim (Yale University), Onkar Dabeer (AWS AI Labs)
RetrievalRepresentation LearningTransformerSupervised Fine-TuningPrompt EngineeringContrastive LearningImage
🎯 What it does: Proposes a visual representation learning setup for open-world continual learning and introduces the DPaRL method for this scenario;
OPEN: Object-wise Position Embedding for Multi-view 3D Object Detection
Jinghua Hou (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
Object DetectionAutonomous DrivingTransformerImagePoint Cloud
🎯 What it does: Proposes a multi-view 3D object detection method called OPEN, which improves detection by leveraging object-level depth information.
OpenIns3D: Snap and Lookup for 3D Open-vocabulary Instance Segmentation
Zhening Huang (University of Cambridge), Joan Lasenby (University of Cambridge)
SegmentationLarge Language ModelVision Language ModelPoint Cloud
🎯 What it does: Proposes the OpenIns3D framework, achieving 3D open-vocabulary instance segmentation without 2D image input.
OpenKD: Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection
Changsheng Lu (University of Science and Technology of China), Piotr Koniusz (University of Technology Sydney)
RecognitionPose EstimationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a facial recognition framework based on pose information fusion, which can extract and integrate multi-directional features under different perspectives to achieve more robust facial matching;
OpenPSG: Open-set Panoptic Scene Graph Generation via Large Multimodal Models
Zijian Zhou (Tongji University), Miaojing Shi (Delft University of Technology)
SegmentationGenerationTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Proposes an open panoramic scene graph generation (OpenPSG) based on large multimodal models, which can simultaneously perform open object segmentation and relation prediction in images.
OpenSight: A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection
Hu Zhang (CSIRO DATA61), Kaicheng Yu (Westlake University)
Object DetectionAutonomous DrivingTransformerLarge Language ModelVision Language ModelImageTextPoint Cloud
🎯 What it does: Designed an open-vocabulary LiDAR object detection framework called OpenSight, which can first identify general objects and then assign specific categories without requiring labeled annotations.
Operational Open-Set Recognition and PostMax Refinement
Steve Cruz (University of Notre Dame), Terrance E. Boult (University of Colorado Colorado Springs)
ClassificationRecognitionConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposed an end-to-end evaluation framework for large-scale open set recognition, and introduced a novel post-processing algorithm called PostMax within this framework to enhance the open set recognition performance of pre-trained closed-set classifiers.
OphNet: A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding
Ming Hu (Monash University), Zongyuan Ge (Monash University)
ClassificationRecognitionConvolutional Neural NetworkTransformerVision Language ModelVideoBenchmark
🎯 What it does: Constructed a large-scale ophthalmic surgery video benchmark called OphNet, and evaluated four tasks on this benchmark: surgery recognition, stage and operation recognition, stage localization, and stage prediction.
Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data
Xiaofan Que (Rochester Institute of Technology), Qi Yu (Rochester Institute of Technology)
Domain AdaptationRepresentation LearningMeta LearningGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Proposes the DUNT (Diverse Unsupervised Tasks for NFSL) framework, which compensates for information loss caused by data cleaning in noisy few-shot learning by leveraging carefully constructed unsupervised auxiliary tasks, and achieves robust representation learning through episode cleansing, contrastive learning, and domain discriminators.
Optimization-based Uncertainty Attribution Via Learning Informative Perturbations
Hanjing Wang (Rensselaer Polytechnic Institute), Qiang Ji (Rensselaer Polytechnic Institute)
SegmentationOptimizationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: By optimizing the learning of binary masks and applying learnable Gaussian blur perturbations, this approach identifies and reduces the sources of uncertainty in deep learning model predictions, achieving uncertainty attribution.
Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation
Yixiao Wang (University of California, Berkeley), Wei Zhan (University of California, Berkeley)
Autonomous DrivingOptimizationComputational EfficiencyDiffusion modelScore-based ModelTime SeriesSequentialStochastic Differential Equation
🎯 What it does: Propose an improved diffusion model for joint trajectory prediction and controllable trajectory generation, significantly reducing inference cost through two methods: Optimal Gaussian Diffusion (OGD) and Estimated Clean Manifold Guidance (ECM);
Optimizing Factorized Encoder Models: Time and Memory Reduction for Scalable and Efficient Action Recognition
Shreyank N Gowda (University of Oxford), Jonathan Huang (Google Research)
RecognitionComputational EfficiencyTransformerSupervised Fine-TuningVideo
🎯 What it does: This paper proposes a two-stage training strategy: first, fully train the spatial and temporal Transformers of ViViT on a small number of frames (8 frames) to initialize the temporal Transformer, then freeze the spatial Transformer and add a lightweight Adapter, fine-tuning only the temporal Transformer and Adapter, thus significantly reducing training time and memory requirements, and enabling the use of a larger spatial backbone.
Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging
Mahmoud Afifi (Google), Liang Liang (Google)
OptimizationComputational EfficiencyRepresentation LearningImage
🎯 What it does: By leveraging dual exposure frames from HDR cameras, we propose Dual Exposure Features (DEF) to enhance the brightness estimator.
Osmosis: RGBD Diffusion Prior for Underwater Image Restoration
Opher Bar Nathan (University of Haifa), Dan Rosenbaum (University of Haifa)
RestorationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Proposes a method for unsupervised underwater image restoration using RGBD diffusion prior and underwater imaging model.
OTSeg: Multi-prompt Sinkhorn Attention for Zero-Shot Semantic Segmentation
Kwanyoung Kim (Korea Advanced Institute of Science and Technology), Jong Chul Ye (Korea Advanced Institute of Science and Technology)
SegmentationTransformerVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes the OTSeg framework, which achieves zero-shot semantic segmentation by utilizing multi-prompt Sinkhorn attention.
Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)
Marko Savic (University of Oulu), Guoying Zhao (University of Oulu)
Adversarial AttackConvolutional Neural NetworkTransformerSupervised Fine-TuningContrastive LearningVideoBiomedical DataBenchmark
🎯 What it does: Constructed and released the first database targeting physical domain attacks on remote photoplethysmography (rPPG) (ORPDAD), evaluating the robustness of various handcrafted features and deep learning methods against three types of attacks (illumination, motion, occlusion).
Out-of-Bounding-Box Triggers: A Stealthy Approach to Cheat Object Detectors
Tao Lin (Chinese Academy of Sciences), Lijun Zhang (Chinese Academy of Sciences)
Object DetectionOptimizationAdversarial AttackImageVideo
🎯 What it does: Propose a stealthy trigger attack outside the target boundary on object detection models, enabling the model to ignore the target.
OV-Uni3DETR: Towards Unified Open-Vocabulary 3D Object Detection via Cycle-Modality Propagation
Zhenyu Wang (Tsinghua University), Shengjin Wang (Tsinghua University)
Object DetectionAutonomous DrivingConvolutional Neural NetworkTransformerImageMultimodalityPoint Cloud
🎯 What it does: Propose a unified open-vocabulary 3D object detection framework, OV-Uni3DETR, which can perform 3D detection under any sensor modality (point cloud, RGB image) and supports multiple indoor and outdoor scenarios;
Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection
Yunfeng FAN, Song Guo (Hong Kong University of Science and Technology)
Federated LearningMultimodality
🎯 What it does: Propose a balanced modal selection framework, BMSFed, which utilizes global prototypes to enhance weak modalities and dynamically selects multi-modal and single-modal clients through sub-mode sub-functions, thereby overcoming modal bias in multi-modal federated learning.
Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks
Cheeun Hong (Seoul National University), Kyoung Mu Lee (Seoul National University)
Super ResolutionImage
🎯 What it does: Propose a novel quantization-aware training framework ODM to address the distribution mismatch problem in SR networks without requiring dynamic quantization ranges.
OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks
jingyang xiang, Yong Liu (Hangzhou Dianzi University)
OptimizationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: This paper investigates the 'silent weights' problem during the weight sign update process in binary neural networks, and proposes the OvSW method to address this issue.
PACE: Pose Annotations in Cluttered Environments
Yang You (Stanford University), Cewu Lu (University Of California San Diego)
Pose EstimationSimultaneous Localization and MappingOptical FlowImageVideoBenchmark
🎯 What it does: This paper proposes a large-scale real and simulated pose annotation dataset called PACE, and analyzes the performance of existing 6D object pose estimation methods in complex occlusion and movable object scenarios through an evaluation benchmark system.
PairingNet: A Learning-based Pair-searching and -matching Network for Image Fragments
Rixin Zhou (Jilin University), chuntao li
RestorationData SynthesisRepresentation LearningGraph Neural NetworkTransformerContrastive LearningImage
🎯 What it does: This paper proposes a deep learning-based image fragment pairing search and matching network called PairingNet, aiming to solve the problem of quickly and accurately finding adjacent fragment pairs in mixed fragment sets.
Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution
Yuehan Zhang (National University of Singapore), Angela Yao (National University of Singapore)
Super ResolutionDomain AdaptationKnowledge DistillationRepresentation LearningImage
🎯 What it does: For unsupervised real-world image super-resolution (RWSR), the authors propose a framework based on pairwise distance distillation (PDD), which leverages existing specialized models (focused on synthetic denoising) and general-purpose models (targeting diverse denoising scenarios). By maintaining consistency in feature distances both within and across models on unannotated real low-resolution images, the framework further enhances the super-resolution quality of specialized models on real images.
PALM: Predicting Actions through Language Models
Sanghwan Kim (ETH Zürich), Xi Wang (ETH Zürich)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoTextMultimodality
🎯 What it does: This paper proposes the PALM framework, which realizes long-term action prediction in first-person videos by combining action recognition, image description, and large language models (LLMs);
Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration
Youngjin Oh (Seoul National University), Nam Ik Cho (Seoul National University)
RestorationTransformerContrastive LearningImage
🎯 What it does: Propose a UDC image restoration framework based on panel-specific degradation representation embedded in Transformer, and collect raw UDC datasets from two commercial phone panels (ZTE Axon 30 5G, Samsung Galaxy Z-Fold 3).
PanGu-Draw: Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion
Guansong Lu (Huawei Noah's Ark Lab), Hang Xu (Huawei Noah's Ark Lab)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelAuto EncoderImageTextMultimodality
🎯 What it does: Propose PanGu-Draw, a resource-efficient text-to-image generation model based on a time-decoupled training strategy and the Coop-Diffusion algorithm.
PanoFree: Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance
Aoming Liu (OPPO US Research Center), Bryan Plummer (Boston University)
GenerationData SynthesisPrompt EngineeringDiffusion modelImageTextStochastic Differential Equation
🎯 What it does: Propose PanoFree, a no-tuning panoramic multi-view image generation method that generates multi-view images using iterative deformation and filling techniques, and supports various correspondence relationships.
PanoVOS: Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation
Shilin Yan (Fudan University), Wei Zhang (Fudan University)
SegmentationTransformerVideo
🎯 What it does: This paper proposes the PanoVOS task for panoramic video object segmentation, constructs a dataset of 150 high-resolution panoramic videos, and designs a Transformer network called PSCFormer aimed at panoramic spatial consistency.
PapMOT: Exploring Adversarial Patch Attack against Multiple Object Tracking
Jiahuan Long (Shanghai Jiao Tong University), Xiaoqian Chen (Chinese Academy of Military Science)
Object TrackingAdversarial AttackVideo
🎯 What it does: Propose PapMOT, a printable physical adversarial patch designed for multi-object tracking (MOT) systems, which can simultaneously interfere with detectors and associators, leading to identity errors or the emergence of new identities.
PaPr: Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference
Tanvir Mahmud (University of Texas at Austin), Diana Marculescu (University of Texas at Austin)
ClassificationRecognitionComputational EfficiencyConvolutional Neural NetworkTransformerImageVideo
🎯 What it does: Propose a lightweight convolutional network (PaPr) scheme for accelerating inference of vision Transformers, convolutional networks, and hybrid Transformers, which is training-agnostic and capable of one-time pruning of image/video patches;
Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer: A Disentangled Approach
Taolin Zhang (Tsinghua University), Shu-Tao Xia (Tsinghua University)
ClassificationComputational EfficiencyTransformerSupervised Fine-TuningImage
🎯 What it does: Proposes a parameter and memory-efficient transfer learning framework called SynQT based on Vision Transformer, which significantly reduces GPU memory usage during training while maintaining high performance.
Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering
Baixin Xu (Nanyang Technological University), Ying He (Nanyang Technological University)
GenerationImage
🎯 What it does: This work proposes an end-to-end neural parametric framework that can automatically learn parameterized domains for spheres or multi-cubes using only 2D multi-view images. It maps neural implicit surfaces to the domain via bidirectional deformation, decomposing the volumetric radiance field into view-independent material and view-dependent shadows, enabling localized editing of texture and material.
Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer
Xueyi Liu (Tsinghua University), Li Yi (Tsinghua University)
OptimizationRobotic IntelligenceSequentialPhysics Related
🎯 What it does: Proposes a parameterized quasi-physical simulator and its physical curriculum for transferring human manipulation demonstrations to robotic hand simulations.
ParCo: Part-Coordinating Text-to-Motion Synthesis
Qiran Zou (Tsinghua University), Xiangyang Ji (Dalian University Of Technology)
GenerationData SynthesisTransformerVision Language ModelAuto EncoderTextMultimodality
🎯 What it does: This paper proposes a text-driven motion generation framework called ParCo, which first decomposes full-body actions into six local parts (left and right hands, left and right feet, spine, root). Each part is discretized and encoded using VQ-VAE; subsequently, six lightweight Transformers generate code sequences for each part, and the Part Coordination module enables communication between generators to synchronously produce coordinated, fine-grained full-body motions.
PARE-Net: Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration
Runzhao Yao (Xi'an Jiaotong University), Chengwu Yang (Xi'an Jiaotong University)
Pose EstimationConvolutional Neural NetworkTransformerContrastive LearningPoint Cloud
🎯 What it does: Designed a lightweight point cloud registration network called PARE-Net based on position-aware rotation equivariant convolution, which can learn rotation equivariant features and achieve efficient and robust registration without relying on rotation data augmentation.
PARIS3D: Reasoning-based 3D Part Segmentation Using Large Multimodal Model
Amrin Kareem (Mohamed Bin Zayed University of Artificial Intelligence), Hisham Cholakkal (Mohamed Bin Zayed University of Artificial Intelligence)
SegmentationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityPoint CloudMesh
🎯 What it does: Propose a 3D reasoning segmentation task based on a multimodal large language model—segmenting parts of 3D objects and generating explanations under natural language implicit queries.
Parrot Captions Teach CLIP to Spot Text
Yiqi Lin (Shanghai Artificial Intelligence Laboratory), Mike Zheng Shou (National University of Singapore)
Data SynthesisExplainability and InterpretabilityRepresentation LearningVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper conducts a systematic study on the bias of the CLIP model in visual-text recognition. It first analyzes the prevalence of 'Parrot Captions' (i.e., captions that heavily overlap with visual text in images) in the LAION-2B dataset. Subsequently, the impact of these captions on model training and downstream tasks is evaluated through experiments involving text removal, synthetic text prompts, and retraining CLIP on different subsets.
Parrot: Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation
Seung Hyun Lee (Google Research), Feng Yang (Google Research)
GenerationTransformerReinforcement LearningPrompt EngineeringDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes Parrot, a multi-objective reinforcement learning framework that jointly fine-tunes the Prompt extension network and the Stable Diffusion generative model, achieving improvements in four metrics: aesthetics, human preference, text-image alignment, and emotional appeal.
Part2Object: Hierarchical Unsupervised 3D Instance Segmentation
Cheng Shi (ShanghaiTech University), Sibei Yang (ShanghaiTech University)
SegmentationConvolutional Neural NetworkTransformerContrastive LearningPoint Cloud
🎯 What it does: Proposes an unsupervised 3D instance segmentation framework called Part2Object based on hierarchical clustering and 2D RGB frame guidance, and trains Hi-Mask3D on this framework to achieve hierarchical object and part segmentation.
PartCraft: Crafting Creative Objects by Parts
Kam Woh Ng, Tao Xiang
Pose EstimationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a deep convolutional network based on a multi-scale attention mechanism for precise detection of facial landmarks.
PartGLEE: A Foundation Model for Recognizing and Parsing Any Objects
Junyi Li (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
Object DetectionSegmentationTransformerVision Language ModelImageText
🎯 What it does: Propose PartGLEE, a hierarchical foundation model based on Q-Former, capable of detecting, segmenting, and localizing objects and parts at any level in an open world.
PartImageNet++ Dataset: Scaling up Part-based Models for Robust Recognition
Xiao Li (Tsinghua University), Xiaolin Hu (Harbin Institute of Technology)
ClassificationRecognitionComputational EfficiencyRepresentation LearningAdversarial AttackData-Centric LearningConvolutional Neural NetworkSupervised Fine-TuningImageBenchmark
🎯 What it does: Constructed the PartImageNet++ (PIN++) dataset and proposed a multi-scale part-supervised model (MPM) based on this dataset to enhance adversarial robustness in image classification and robustness to common noise and out-of-distribution (OOD) samples.
PartSTAD: 2D-to-3D Part Segmentation Task Adaptation
Hyunjin Kim (KRAFTON Inc.), Minhyuk Sung (KAIST)
SegmentationDomain AdaptationTransformerPoint Cloud
🎯 What it does: Propose the PartSTAD method, leveraging GLIP's 2D detection and SAM's masks, using a lightweight MLP to predict boundary weights, achieving 3D part segmentation with few samples;
PatchRefiner: Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation
Zhenyu Li (King Abdullah University of Science and Technology), Peter Wonka (King Abdullah University of Science and Technology)
Depth EstimationDomain AdaptationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: Propose the PatchRefiner framework, transforming high-resolution monocular depth estimation into a refinement process of rough depth, and enhance real-world performance by leveraging synthetic data pseudo-labels through a teacher-student structure.
Pathformer3D: A 3D Scanpath Transformer for 360° Images
Rong Quan (Nanjing University of Aeronautics and Astronautics), Dong Liang (Nanjing University of Aeronautics and Astronautics)
TransformerGaussian SplattingImage
🎯 What it does: Predict the scan path of 360° images using a 3D spherical coordinate system and Transformer to generate continuous fixation points.
PathMMU: A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology
Yuxuan Sun (Zhejiang University), Lin Yang (Westlake University)
Convolutional Neural NetworkLarge Language ModelSupervised Fine-TuningImageTextMultimodalityBiomedical DataBenchmark
🎯 What it does: This study proposes PathMMU—a largest-scale, highest-quality, expert-verified multimodal pathology benchmark containing 33,428 multiple-choice questions and 24,067 pathology images, each with detailed answer explanations. During construction, GPT-4V was first used to enhance descriptions of original image-text pairs, followed by Q&A generation. Questions were then filtered through multiple rounds of LLM screening and reviewed by seven pathology experts to ensure both image analysis requirements and professional rigor. Subsequently, 18 advanced multimodal models (including open-source and closed-source) were evaluated on PathMMU via zero-shot assessment, tested for robustness under image distortion, and fine-tuned to evaluate transfer learning capabilities.
Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification
Linhao Qu (Shanghai Artificial Intelligence Laboratory), Xiaosong Wang (Shanghai Artificial Intelligence Laboratory)
ClassificationTransformerPrompt EngineeringVision Language ModelContrastive LearningMultimodalityBiomedical Data
🎯 What it does: Propose a multi-instance prompt learning framework PEMP that leverages pathological visual and text priors to enhance performance in few-shot scenarios for whole slide image (WSI) classification.
PAV: Personalized Head Avatar from Unstructured Video Collection
Akin Caliskan (Flawless AI Imperial College London), Hyeongwoo Kim
GenerationPose EstimationNeural Radiance FieldVideo
🎯 What it does: Proposed PAV, a method capable of learning and rendering personalized dynamic head neural radiance fields from multi-view, unstructured video collections;
Paying More Attention to Images: A Training-Free Method for Alleviating Hallucination in LVLMs
Shi Liu (Zhejiang University), Wei Chen (Zhejiang University)
OptimizationTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Propose a training-agnostic inference intervention method called PAI, dynamically amplifying visual attention and reducing text priors to mitigate hallucinations in large vision-language models.
PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion
Runsong Zhu (Chinese University of Hong Kong), Chi-Wing Fu (Chinese University of Hong Kong)
SegmentationNeural Radiance FieldContrastive LearningImage
🎯 What it does: Designed a Panoptic Lifting framework called PCF-Lift based on probability contrastive fusion, which projects multi-view 2D panoptic segmentation into 3D space and generates consistent 3D panoptic segmentation results.
PDiscoFormer: Relaxing Part Discovery Constraints with Vision Transformers
Ananthu Aniraj (Inria), Diego Marcos (Inria)
SegmentationRepresentation LearningTransformerImage
🎯 What it does: Proposes an unsupervised part discovery framework called PDiscoFormer based on Vision Transformer, which utilizes total variation prior to relax constraints on part size and shape, achieving automatic localization of semantic parts in fine-grained images.
PDT Uav Target Detection Dataset for Pests and Diseases Tree
Mingle Zhou (Qilu University of Technology (Shandong Academy of Sciences)), Gang Li (Qilu University of Technology (Shandong Academy of Sciences))
Object DetectionConvolutional Neural NetworkImageAgriculture Related
🎯 What it does: Constructed a high-resolution and low-resolution image dataset for UAVs (PDT) and integrated public data to generate a multi-class crop and weed dataset (CWC). Subsequently, a specialized YOLO-DP model for dense small target detection was designed based on YOLOv5s; compared the performance of existing detection models on three datasets.
PEA-Diffusion: Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation
jian ma, Haonan Lu (OPPO AI Center)
GenerationKnowledge DistillationConvolutional Neural NetworkTransformerVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Propose a cross-lingual text-to-image generation framework PEA‑Diffusion that combines a lightweight parameter-efficient adapter (PEA) with knowledge distillation, enabling the transfer of Stable Diffusion's generative capabilities to non-English cultural contexts without training the UNet.
Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting
Jeongmin Bae (Yonsei University), Youngjung Uh (Yonsei University)
GenerationGaussian SplattingVideo
🎯 What it does: Proposes a deformation method based on latent embeddings for each Gaussian to improve the rendering quality of deformable 3D Gaussian Splatting in complex dynamic scenes.
Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers’ Opinion Scores
Lucas Goncalves (University of Texas at Dallas), Kyu Han
Explainability and InterpretabilityRepresentation LearningData-Centric LearningConvolutional Neural NetworkTransformerContrastive LearningVideoMultimodalityBenchmarkAudio
🎯 What it does: Constructed an audio-visual synchronization evaluation dataset covering nine types of synchronization errors and over 100 hours of human-annotated audio-visual data, and proposed an interpretable 5-point automatic evaluation metric called PEAVS to measure the audio-visual synchronization quality of videos.
Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching
Yichen Li (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
Domain AdaptationFederated LearningKnowledge DistillationImage
🎯 What it does: Proposes a personalized approach called pFedDIL for federated domain incremental learning (FDIL), which can gradually learn new tasks while retaining knowledge of old tasks in multi-client, multi-domain scenarios.
Personalized Privacy Protection Mask Against Unauthorized Facial Recognition
Ka-Ho Chow (University of Hong Kong), Ling Liu (Georgia Institute of Technology)
RecognitionSafty and PrivacyImage
🎯 What it does: Propose a personalized privacy protection mask called P3-Mask for each user to instantly protect user facial photos from unauthorized face recognition without requiring manual per-image optimization.
Personalized Video Relighting With an At-Home Light Stage
Jun Myeong Choi (University of North Carolina at Chapel Hill), Roni Sengupta (University of North Carolina at Chapel Hill)
Image TranslationRestorationConvolutional Neural NetworkGenerative Adversarial NetworkVideo
🎯 What it does: This paper proposes a personalized video relighting algorithm that utilizes a user's home computer monitor as a light source, capable of performing high-quality and temporally consistent relighting on videos with arbitrary poses, expressions, and environmental lighting conditions in real-time (≈45fps).
PetFace: A Large-Scale Dataset and Benchmark for Animal Identification
Risa Shinoda (Kyoto University), Kaede Shiohara (University of Tokyo)
RecognitionConvolutional Neural NetworkContrastive LearningImageBenchmark
🎯 What it does: This paper proposes a large-scale animal facial recognition dataset called PetFace, and designs benchmarks for two tasks: re-identification of known individuals and verification of unknown individuals, including data collection, automatic alignment, manual filtering, and fine-grained annotation.
PFedEdit: Personalized Federated Learning via Automated Model Editing
Haolin Yuan (Johns Hopkins University), Yinzhi Cao (Johns Hopkins University)
Federated LearningImageBiomedical DataBenchmark
🎯 What it does: Propose the PFedEdit framework, which automatically selects personalized layers through model editing to enhance model performance in non-i.i.d. environments within federated learning.
PFGS: High Fidelity Point Cloud Rendering via Feature Splatting
Jiaxu Wang (Hong Kong University of Science and Technology), Renjing Xu (Hong Kong University of Science and Technology)
GenerationConvolutional Neural NetworkGaussian SplattingPoint Cloud
🎯 What it does: Integrate sparse point cloud rendering with 3D Gaussian Splatting to propose the PFGS framework, achieving high-fidelity, 3D-consistent image rendering.
Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation
Hoyong Kwon (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
SegmentationTransformerImageBenchmark
🎯 What it does: Proposes two modules based on the frequency domain: Magnitude-mixing-based Phase Concentration (MPC) and Frequency Shortcut Suppression (FSS), to enhance the Class Activation Map (CAM) quality of Vision Transformers (ViT) in weakly supervised semantic segmentation, reducing boundary blurriness and erroneous activation.
Photon Inhibition for Energy-Efficient Single-Photon Imaging
Lucas J Koerner (University of St Thomas), Mohit Gupta (University of Wisconsin Madison)
Computational EfficiencyImagePhysics Related
🎯 What it does: Propose a power consumption control method for single-photon cameras based on adaptive suppression, which real-time shuts down part of the SPAD pixels to reduce photon detection quantity