ICCV 2023 Papers — Page 14
IEEE/CVF International Conference on Computer Vision · 2156 papers
Neural Haircut: Prior-Guided Strand-Based Hair Reconstruction
Vanessa Sklyarova (Samsung AI Center), Egor Zakharov (Samsung AI Center)
RestorationGenerationNeural Radiance FieldImageVideo
🎯 What it does: This paper proposes a two-stage image/video-driven hair reconstruction method, which first reconstructs rough hair and facial geometry through implicit volume reconstruction, and then generates a high-quality single-strand level 3D hair model using prior-based fine-grained strand optimization.
Neural Implicit Surface Evolution
Tiago Novello, Luiz Velho
Point CloudMesh
🎯 What it does: Utilizing smooth neural networks to represent and learn implicit surfaces in the spatiotemporal domain (R³×R) under the continuous evolution driven by the level set equation (LSE); training is achieved using only the initial SDF conditions and constrained by an unsupervised LSE loss.
Neural Interactive Keypoint Detection
Jie Yang (International Digital Economy Academy), Lei Zhang (International Digital Economy Academy)
Pose EstimationTransformerImage
🎯 What it does: An interactive 2D human keypoint detection framework called Click-Pose is proposed, where users can correct predicted keypoints with minimal clicks, and the remaining keypoints are automatically optimized, supporting end-to-end inference without the need for post-processing.
Neural LiDAR Fields for Novel View Synthesis
Shengyu Huang (ETH Zurich), Or Litany (NVIDIA)
GenerationData SynthesisAutonomous DrivingNeural Radiance FieldPoint Cloud
🎯 What it does: This paper studies a neural field method called NFL, which can learn implicit scene representations from LiDAR scans and synthesize realistic LiDAR scans from new viewpoints.
Neural Microfacet Fields for Inverse Rendering
Alexander Mai (University of California San Diego), Sara Fridovich-Keil (University of California Berkeley)
RestorationData SynthesisOptimizationNeural Radiance FieldImage
🎯 What it does: This paper proposes the Neural Microfacet Fields method, which combines volume rendering with microfacet Bidirectional Reflectance Distribution Function (BRDF) to simultaneously recover scene geometry, materials, and environmental lighting from calibrated images.
Neural Radiance Field with LiDAR maps
MingFang Chang, Simon Lucey (University of Adelaide)
RestorationGenerationData SynthesisAutonomous DrivingNeural Radiance FieldGenerative Adversarial NetworkImagePoint Cloud
🎯 What it does: A point-based NeRF framework utilizing LiDAR maps and conditional GANs is designed to generate novel viewpoint images from outdoor camera-LiDAR datasets.
Neural Reconstruction of Relightable Human Model from Monocular Video
Wenzhang Sun (Beijing Institute of Technology), Yandong Guo (AI 2 Robotics)
GenerationPose EstimationNeural Radiance FieldImageVideo
🎯 What it does: A self-supervised framework is proposed that can reconstruct re-illuminable and animatable 3D human models from monocular videos and render them under arbitrary poses and lighting conditions.
Neural Video Depth Stabilizer
Yiran Wang (Huazhong University of Science and Technology), Guosheng Lin (Nanyang Technological University)
RestorationDepth EstimationTransformerOptical FlowVideo
🎯 What it does: A pluggable video depth stabilizer NVDS is proposed, which can eliminate inter-frame flicker and enhance spatiotemporal consistency without modifying the single-image depth model; a large-scale natural scene video depth dataset VDW is also released.
Neural-PBIR Reconstruction of Shape, Material, and Illumination
Cheng Sun (Meta), Zhao Dong (Meta)
RestorationOptimizationKnowledge DistillationNeural Radiance FieldImage
🎯 What it does: A three-stage pipeline is proposed, combining neural SDF with physics-based inverse rendering to achieve high-quality recovery of object geometry, materials, and environmental lighting, while supporting re-rendering.
NeuRBF: A Neural Fields Representation with Adaptive Radial Basis Functions
Zhang Chen (OPPO US Research Center), Yi Xu (University at Buffalo)
GenerationData SynthesisKnowledge DistillationRepresentation LearningNeural Radiance FieldImagePoint Cloud
🎯 What it does: A neural field representation NeuRBF based on adaptive radial basis functions (RBF) is designed and implemented, combining grid-based interpolation with adaptive RBF, and employing multi-frequency sine expansion to enhance representation capability. Its effectiveness is validated in 2D image fitting, 3D SDF reconstruction, and NeRF generation tasks.
NeuS2: Fast Learning of Neural Implicit Surfaces for Multi-view Reconstruction
Yiming Wang (University of Pennsylvania), Lingjie Liu (Max Planck Institute for Informatics)
GenerationComputational EfficiencyNeural Radiance FieldVideoPoint Cloud
🎯 What it does: We propose NeuS2, a neural implicit surface reconstruction method based on multi-resolution hash encoding, capable of high-quality reconstruction of static scenes in a few minutes and dynamic scenes in about 20 seconds per frame.
NIR-assisted Video Enhancement via Unpaired 24-hour Data
Muyao Niu (University of Tokyo), Yinqiang Zheng (University of Tokyo)
Image TranslationRestorationData SynthesisTransformerGenerative Adversarial NetworkVideo
🎯 What it does: Using unpaired all-weather (24-hour) visible and near-infrared video for low-light video enhancement, an end-to-end model based on physics-inspired illumination redirection, noise GAN, and temporal perception network is proposed.
NLOS-NeuS: Non-line-of-sight Neural Implicit Surface
Yuki Fujimura (Nara Institute of Science and Technology), Yasuhiro Mukaigawa (Nara Institute of Science and Technology)
GenerationDepth EstimationNeural Radiance FieldPoint CloudMesh
🎯 What it does: A new method called NLOS-NeuS is proposed for reconstructing 3D surfaces in non-line-of-sight imaging scenarios, extending the Neural Transient Field (NeTF) to support neural implicit surface representation.
No Fear of Classifier Biases: Neural Collapse Inspired Federated Learning with Synthetic and Fixed Classifier
Zexi Li (Zhejiang University), Chao Wu (Zhejiang University)
Federated LearningImage
🎯 What it does: Proposes the use of a fixed Equiangular Tight Frame (ETF) classifier in federated learning to eliminate classifier bias caused by client data heterogeneity, and achieves personalization by alternately fine-tuning the ETF classifier and the projection layer after training.
Noise-Aware Learning from Web-Crawled Image-Text Data for Image Captioning
Wooyoung Kang (Kakao Brain), Byungseok Roh (Kakao Brain)
GenerationRetrievalTransformerContrastive LearningImageTextMultimodality
🎯 What it does: A noise-aware learning framework called NoC is proposed, which utilizes alignment level control of the generator to train image description models on the crawled image-text data from the entire web.
Noise2Info: Noisy Image to Information of Noise for Self-Supervised Image Denoising
Jiachuan Wang (Hong Kong University of Science and Technology), Charles Wang Wai Ng (Hong Kong University of Science and Technology)
RestorationConvolutional Neural NetworkImage
🎯 What it does: The Noise2Info method is proposed, which dynamically estimates the noise standard deviation and performs image denoising using a self-supervised framework with only a single noisy image.
Non-Coaxial Event-Guided Motion Deblurring with Spatial Alignment
Hoonhee Cho (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
RestorationTransformerImageMultimodalityBenchmark
🎯 What it does: A motion deblurring method called NED-Net is proposed for the combination of non-coaxial event cameras and RGB cameras, which can recover clear images using event information without the need for pixel-level alignment.
Non-Semantics Suppressed Mask Learning for Unsupervised Video Semantic Compression
Yuan Tian (Shanghai Jiao Tong University), Zhiyong Gao (Shanghai Jiao Tong University)
RecognitionObject TrackingSegmentationCompressionAuto EncoderGenerative Adversarial NetworkVideo
🎯 What it does: This paper proposes an unsupervised video semantic compression framework (SMC) that first extracts the semantic features of the original video and then compensates for the semantic loss in the compressed video, achieving semantically complete video coding with a low compression rate.
Nonrigid Object Contact Estimation With Regional Unwrapping Transformer
Wei Xie (Southeast University), Yangang Wang (Southeast University)
Object DetectionPose EstimationTransformerImagePoint Cloud
🎯 What it does: A learning framework is proposed to estimate the contact patterns and deformations between the palm and non-rigid objects from a monocular image, utilizing the hand-object surface area unfolded into fine-grained 2D images to predict contact and deformation.
Normalizing Flows for Human Pose Anomaly Detection
Or Hirschorn (Tel Aviv University), Shai Avidan (Tel Aviv University)
Pose EstimationAnomaly DetectionGraph Neural NetworkFlow-based ModelVideoTime Series
🎯 What it does: This paper proposes a lightweight anomaly detection framework STG-NF that utilizes only human skeletal sequences, based on regularized flow to learn skeletal distribution and assess anomaly probability through log-likelihood.
Not All Features Matter: Enhancing Few-shot CLIP with Adaptive Prior Refinement
Xiangyang Zhu (City University of Hong Kong), Peng Gao (Chinese University of Hong Kong)
ClassificationRecognitionConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Enhancing few-shot learning for CLIP, the APE method is proposed to improve CLIP's performance in few-shot classification tasks through adaptive prior refinement.
Not All Steps are Created Equal: Selective Diffusion Distillation for Image Manipulation
Luozhou Wang (Hong Kong University of Science and Technology), Ying-cong Chen (Hong Kong University of Science and Technology)
GenerationData SynthesisDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: A Selective Diffusion Distillation (SDD) framework is proposed, which uses a pre-trained diffusion model to guide lightweight image manipulators (such as the mapping network of StyleGAN) for image editing through a single forward inference, avoiding the trade-off between editability and fidelity in traditional diffusion editing.
Not Every Side Is Equal: Localization Uncertainty Estimation for Semi-Supervised 3D Object Detection
Chuxin Wang (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
Object DetectionAutonomous DrivingPoint Cloud
🎯 What it does: This study proposes a semi-supervised 3D object detection framework based on side perception, utilizing the probability distribution and geometric features of each side to estimate uncertainty, thereby performing weighted filtering and training of pseudo-labels.
Novel Scenes & Classes: Towards Adaptive Open-set Object Detection
Wuyang Li (City University of Hong Kong), Yixuan Yuan (The Chinese University of Hong Kong)
Object DetectionDomain AdaptationTransformerImage
🎯 What it does: This paper studies Adaptive Open Set Object Detection (AOOD), which considers new scenes in the target domain while also addressing the detection of new categories.
Novel-View Synthesis and Pose Estimation for Hand-Object Interaction from Sparse Views
Wentian Qu (Institute of Software, Chinese Academy of Sciences), Hongan Wang (Institute of Software, Chinese Academy of Sciences)
Pose EstimationNeural Radiance FieldImage
🎯 What it does: This paper proposes a two-stage hand-object interaction neural rendering and pose estimation system. In the offline stage, an implicit model of the hand and object shapes and appearances is trained using sparse views. In the online stage, a rendering-based joint model is fitted with geometric constraints and a stable contact loss to achieve hand-object interaction pose estimation and free-view rendering under sparse views.
NPC: Neural Point Characters from Video
Shih-Yang Su (University of British Columbia), Helge Rhodin (Reality Labs Research)
GenerationPose EstimationGraph Neural NetworkNeural Radiance FieldVideoPoint Cloud
🎯 What it does: A neural point character (NPC) model based on sparse point clouds is proposed, which can learn animatable high-fidelity 3D characters from a single video without the need for pre-existing templates or laser scanning.
NSF: Neural Surface Fields for Human Modeling from Monocular Depth
Yuxuan Xue (University of Tübingen), Tony Tung (Meta)
GenerationData SynthesisPose EstimationDepth EstimationComputational EfficiencyPoint CloudMesh
🎯 What it does: Learning to generate animatable, detail-rich human models from monocular depth sequences.
Object as Query: Lifting Any 2D Object Detector to 3D Detection
Zitian Wang (Institute of Artificial Intelligence), Si Liu (Institute of Artificial Intelligence)
Object DetectionAutonomous DrivingTransformerImage
🎯 What it does: Dynamic queries are generated using a 2D detector and sparse cross-attention with multi-view images to complete 3D object detection.
Object-aware Gaze Target Detection
Francesco Tonini (University of Trento), Elisa Ricci (University of Trento)
Object DetectionTransformerImageVideo
🎯 What it does: A Transformer-based gaze target detection framework is proposed, which can simultaneously predict gaze points, heatmaps, gaze object categories and locations without relying on manually annotated head boxes, and determine whether the gaze falls outside the frame.
Object-Centric Multiple Object Tracking
Zixu Zhao (Amazon Web Services), Tianjun Xiao (Amazon Web Services)
Object TrackingVideo
🎯 What it does: This paper proposes an unsupervised multi-object tracking framework OC-MOT based on video object centric learning, utilizing a self-supervised memory module and an index-merge mechanism to address the issues of partial-whole segmentation and temporal consistency of objects.
ObjectFusion: Multi-modal 3D Object Detection with Object-Centric Fusion
Qi Cai (University of Science and Technology of China), Tao Mei (HiDream.ai Inc)
Object DetectionAutonomous DrivingTransformerMultimodalityPoint Cloud
🎯 What it does: This paper studies a new multimodal 3D object detection framework called ObjectFusion, which utilizes object-centered features to avoid the conversion from camera to BEV, achieving precise fusion of multi-sensor information.
Objects Do Not Disappear: Video Object Detection by Single-Frame Object Location Anticipation
Xin Liu (Delft University of Technology), Silvia L. Pintea (University of Amsterdam)
Object DetectionObject TrackingComputational EfficiencyConvolutional Neural NetworkOptical FlowVideo
🎯 What it does: This paper proposes a scheme that utilizes only key frames for object detection and predicting the target position in subsequent frames, which improves detection accuracy while significantly reducing computational and labeling costs.
ObjectSDF++: Improved Object-Compositional Neural Implicit Surfaces
Qianyi Wu (Monash University), Jianfei Cai (Monash University)
Object DetectionSegmentationPoint Cloud
🎯 What it does: This paper proposes ObjectSDF++, a framework for object-level neural implicit surface reconstruction aimed at multi-view RGB + instance segmentation.
Occ^2Net: Robust Image Matching Based on 3D Occupancy Estimation for Occluded Regions
Miao Fan (MEGVII Technology), Shuchang Zhou (MEGVII Technology)
Pose EstimationImage
🎯 What it does: A framework for image matching called Occ 2 Net is proposed, which can simultaneously match visible points and occluded points in occluded scenes, thereby improving pose estimation accuracy.
OccFormer: Dual-path Transformer for Vision-based 3D Semantic Occupancy Prediction
Yunpeng Zhang (PhiGent Robotics), Dalong Du (PhiGent Robotics)
SegmentationAutonomous DrivingTransformerImagePoint Cloud
🎯 What it does: A dual-path Transformer-based OccFormer network is proposed to predict the 3D semantic occupancy of the environment around vehicles using monocular or multi-camera images.
OCHID-Fi: Occlusion-Robust Hand Pose Estimation in 3D via RF-Vision
Shujie Zhang (Nanyang Technological University), Jun Luo (Nanyang Technological University)
Pose EstimationDomain AdaptationKnowledge DistillationConvolutional Neural NetworkMultimodality
🎯 What it does: By utilizing radio frequency vision (RF-vision) technology, OCHID-Fi achieves 3D hand pose estimation in occluded scenes.
OFVL-MS: Once for Visual Localization across Multiple Indoor Scenes
Tao Xie (Harbin Institute of Technology), Ruifeng Li (Harbin Institute of Technology)
Pose EstimationConvolutional Neural NetworkSimultaneous Localization and MappingImage
🎯 What it does: This paper studies a unified visual localization framework based on multi-task learning, capable of predicting camera poses in multiple indoor scenes at once.
Omnidirectional Information Gathering for Knowledge Transfer-Based Audio-Visual Navigation
Jinyu Chen (Beihang University), Yi Yang (Chinese University of Hong Kong)
Knowledge DistillationRobotic IntelligenceRecurrent Neural NetworkReinforcement LearningAgentic AIMultimodalityAudio
🎯 What it does: The ORAN model is proposed for the audio-visual navigation task, enhancing navigation capabilities by combining cross-task policy distillation (CCPD) and omnidirectional information gathering (OIG).
OmniLabel: A Challenging Benchmark for Language-Based Object Detection
Samuel Schulter (NEC Laboratories America), Dimitris Metaxas (Rutgers University)
Object DetectionTransformerVision Language ModelContrastive LearningImageTextBenchmark
🎯 What it does: The OmniLabel benchmark is proposed, unifying standard detection, open vocabulary detection, and referential expression tasks, providing free text descriptions and negative samples;
OmnimatteRF: Robust Omnimatte with 3D Background Modeling
Geng Lin (University of Maryland), Ayush Saraf (Meta)
SegmentationGenerationData SynthesisConvolutional Neural NetworkNeural Radiance FieldOptical FlowImageVideo
🎯 What it does: This paper proposes a film layering method called OmnimatteRF, which combines a 2D foreground layer with a 3D background voxel neural radiance field. It can simultaneously extract foreground objects along with their shadows, reflections, and other accompanying effects, while generating a clean 3D background.
OmniZoomer: Learning to Move and Zoom in on Sphere at High-Resolution
Zidong Cao (Hong Kong University of Science and Technology), Lin Wang (Hong Kong University of Science and Technology)
Image TranslationRestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: Proposes the OmniZoomer network, which enables movement and zooming of panoramic images, outputting high-resolution, distortion-free results.
On the Audio-visual Synchronization for Lip-to-Speech Synthesis
Zhe Niu (Hong Kong University of Science and Technology), Brian Mak (Hong Kong University of Science and Technology)
GenerationData SynthesisConvolutional Neural NetworkTransformerGenerative Adversarial NetworkVideoMultimodalityAudio
🎯 What it does: A synchronous lip-to-speech (SLTS) model is proposed, and a time-aligned frontend is designed to address the issues of data and model asynchrony, achieving end-to-end speech synthesis.
On the Effectiveness of Spectral Discriminators for Perceptual Quality Improvement
Xin Luo (University of Science and Technology of China), Dong Liu (University of Science and Technology of China)
RestorationSuper ResolutionTransformerGenerative Adversarial NetworkImage
🎯 What it does: This paper studies the effect of spectral discriminators in GAN-based super-resolution and proposes the Spectral Transformer, which combines frequency domain information with Transformers, and the Dual Transformer, which integrates spatial and frequency domain discriminators, thereby improving super-resolution quality and no-reference image quality assessment.
On the Robustness of Normalizing Flows for Inverse Problems in Imaging
Seongmin Hong (Seoul National University), Se Young Chun (Seoul National University)
RestorationSuper ResolutionFlow-based ModelImage
🎯 What it does: This study investigates the robustness of Conditional Normalizing Flow in image inverse problems, revealing and addressing severe artifacts caused by exploding inverses due to out-of-distribution (OOD) conditional inputs.
On the Robustness of Open-World Test-Time Training: Self-Training with Dynamic Prototype Expansion
Yushu Li (South China University of Technology), Kui Jia (South China University of Technology)
ClassificationDomain AdaptationImage
🎯 What it does: This paper proposes a test-time training method aimed at open-world scenarios (including strong out-of-distribution samples), designed to maintain good performance even when strong OOD (unknown category) samples are present in the target domain.
Once Detected, Never Lost: Surpassing Human Performance in Offline LiDAR based 3D Object Detection
Lue Fan (Institute of Automation, Chinese Academy of Sciences), Zhaoxiang Zhang (Institute of Automation, Chinese Academy of Sciences)
Object DetectionObject TrackingAutonomous DrivingPoint Cloud
🎯 What it does: A novel offline LiDAR 3D object detection system called CTRL has been designed, adopting a trajectory-centric detection approach. It first generates initial boxes using a basic detector, then expands trajectories through bidirectional tracking and jointly optimizes the detection results of all frames using a trajectory-level learning module.
One-bit Flip is All You Need: When Bit-flip Attack Meets Model Training
Jianshuo Dong (Tsinghua University), Shu-Tao Xia (Tsinghua University)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a Training-Assisted Bit Flip Attack (TBA), which constructs a high-risk model during the training phase, allowing for the implantation of malicious functionality by flipping only a few bits after deployment.
One-Shot Generative Domain Adaptation
Ceyuan Yang (Shanghai AI Laboratory), Bolei Zhou (University of California, Los Angeles)
GenerationDomain AdaptationGenerative Adversarial NetworkImage
🎯 What it does: Single image generation domain adaptation based on pre-trained GAN, achieving high-quality and diverse image generation for the target domain using only one reference image.
One-shot Implicit Animatable Avatars with Model-based Priors
Yangyi Huang (Zhejiang University), Deng Cai (Zhejiang University)
GenerationPose EstimationNeural Radiance FieldImage
🎯 What it does: Using SMPL geometric priors and CLIP visual-semantic priors, we achieve the generation of one-shot animatable human avatars from a single image through NeRF.
One-Shot Recognition of Any Material Anywhere Using Contrastive Learning with Physics-Based Rendering
Manuel S. Drehwald (Karlsruhe Institute of Technology), Alan Aspuru-Guzik (University of Toronto)
ClassificationRecognitionConvolutional Neural NetworkContrastive LearningImageBenchmark
🎯 What it does: A MatSim dataset and benchmark were proposed, and a Siamese network based on contrastive learning was trained to achieve one-shot recognition of any material state.
Online Class Incremental Learning on Stochastic Blurry Task Boundary via Mask and Visual Prompt Tuning
Jun-Yeong Moon (Kyung Hee University), Gyeong-Moon Park (Kyung Hee University)
ClassificationRecognitionTransformerPrompt EngineeringContrastive LearningImage
🎯 What it does: A new online continuous learning scenario, Si-Blurry, is proposed, and the Mask and Visual Prompt Tuning (MVP) method is designed to address the issues of task forgetting and class imbalance in this scenario.
Online Clustered Codebook
Chuanxia Zheng (University of Oxford), Andrea Vedaldi (University of Oxford)
GenerationData SynthesisAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: This paper proposes an online clustering vector quantization method called CVQ-VAE, which addresses the VQ codebook collapse problem using dynamic initialization and running average updates.
Online Continual Learning on Hierarchical Label Expansion
Byung Hyun Lee (Seoul National University), Se Young Chun (Seoul National University)
ClassificationRecognitionConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a new online continual learning framework—Hierarchical Label Expansion (HLE), which allows the model to gradually expand from coarse-grained categories to fine-grained categories without knowing the task boundaries, simulating the knowledge accumulation process in reality.
Online Prototype Learning for Online Continual Learning
Yujie Wei (Fudan University), Hongming Shan (Fudan University)
ClassificationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A continuous learning framework based on online prototype learning (OnPro) is proposed, which simultaneously learns new and old categories and enhances discriminability through online prototype balancing and adaptive prototype feedback.
OnlineRefer: A Simple Online Baseline for Referring Video Object Segmentation
Dongming Wu (Beijing Institute of Technology), Jianbing Shen (University of Macau)
Object DetectionSegmentationTransformerVideoText
🎯 What it does: An online query propagation framework called OnlineRefer is proposed for real-time segmentation of target objects from videos based on natural language instructions.
Open Set Video HOI detection from Action-Centric Chain-of-Look Prompting
Nan Xi (State University of New York at Buffalo), Junsong Yuan (State University of New York at Buffalo)
RecognitionObject DetectionGraph Neural NetworkTransformerPrompt EngineeringVision Language ModelVideo
🎯 What it does: This paper proposes an action-centered open-domain video human-object interaction (HOI) detection method called ACoLP.
Open-domain Visual Entity Recognition: Towards Recognizing Millions of Wikipedia Entities
Hexiang Hu, Ming-Wei Chang
RecognitionRetrievalTransformerVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes the Open Domain Visual Entity Recognition (OVEN) task and constructs the OVEN-Wiki dataset to evaluate the recognition capabilities of existing large-scale multimodal pre-trained models on millions of Wikipedia entities.
Open-Vocabulary Object Detection With an Open Corpus
Jiong Wang (Zhejiang University), Zhou Zhao (Zhejiang University)
Object DetectionContrastive LearningImageText
🎯 What it does: In open vocabulary object detection (OVD), this paper proposes two key modules by constructing a large-scale 'open corpus' and embedding it into the detection framework: 1) General Objectness Assessment based on the Open Corpus (GOAT), which determines the objectness of candidate boxes using the cluster centers aligned with text and visual features; 2) Category Expansion based on the Open Corpus (CE), which includes an Open Corpus Classifier (OCC) during the detection phase and Negative Clustering Expansion (NCE) during the image-text alignment phase, thereby expanding more open category information in the positive and negative sample space.
Open-vocabulary Object Segmentation with Diffusion Models
Ziyi Li (Shanghai Jiao Tong University), Weidi Xie (Shanghai AI Laboratory)
Object DetectionSegmentationGenerationTransformerDiffusion modelImageText
🎯 What it does: This paper proposes an open-source vocabulary object segmentation method based on Stable Diffusion, which can obtain segmentation masks corresponding to the objects described in the text prompts while generating images.
Open-vocabulary Panoptic Segmentation with Embedding Modulation
Xi Chen (University of Hong Kong), Hengshuang Zhao (University of Hong Kong)
Object DetectionSegmentationTransformerVision Language ModelImage
🎯 What it does: A panoramic segmentation framework OPSNet is proposed for open vocabulary, capable of handling both known and unknown objects.
Open-Vocabulary Semantic Segmentation with Decoupled One-Pass Network
Cong Han (Meituan Inc), Lin Ma (Meituan Inc)
SegmentationTransformerVision Language ModelImage
🎯 What it does: An efficient open vocabulary semantic segmentation network, DeOP, is proposed, which completes segmentation with just one forward pass of the vision-language model.
Open-vocabulary Video Question Answering: A New Benchmark for Evaluating the Generalizability of Video Question Answering Models
Dohwan Ko (Korea University), Hyunwoo J. Kim (Korea University)
RecognitionRetrievalGraph Neural NetworkVideoMultimodalityBenchmark
🎯 What it does: Proposes the Open Vocabulary Video Question Answering (OVQA) benchmark and improves existing VideoQA models on this benchmark to support the prediction of rare and unseen answers.
OpenOccupancy: A Large Scale Benchmark for Surrounding Semantic Occupancy Perception
Xiaofeng Wang (Institute of Automation, Chinese Academy of Sciences), Xingang Wang (Institute of Automation, Chinese Academy of Sciences)
Object DetectionSegmentationAutonomous DrivingComputational EfficiencyMultimodalityPoint CloudBenchmark
🎯 What it does: Proposes the OpenOccupancy evaluation benchmark, extending nuScenes for a surround semantic occupancy perception task, constructing dense semantic occupancy annotations and providing multimodal baselines and a Cascade Occupancy Network.
OPERA: Omni-Supervised Representation Learning with Hierarchical Supervisions
Chengkun Wang (Tsinghua University), Jiwen Lu (Tsinghua University)
ClassificationObject DetectionSegmentationRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: A fully supervised and self-supervised hierarchical supervision joint learning framework named OPERA is designed and implemented to simultaneously utilize both supervised and self-supervised information from labeled and unlabeled data in image representation learning.
Optimizing the Placement of Roadside LiDARs for Autonomous Driving
Wentao Jiang (Beihang University), Si Liu (Shanghai AI Laboratory)
Autonomous DrivingOptimizationPoint Cloud
🎯 What it does: This paper proposes a greedy search algorithm based on perceptual gain and a single-frame perceptual predictor to optimize the placement of roadside LiDAR, thereby enhancing multi-agent collaborative perception performance, and constructs the Roadside-Opt dataset for research.
ORC: Network Group-based Knowledge Distillation using Online Role Change
Junyong Choi (Hyundai Motor Company), Wonjun Hwang (Ajou University)
ClassificationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A knowledge distillation method based on network groups is proposed, employing an Online Role Change (ORC) mechanism that dynamically elevates the best-performing student networks to temporary teachers during the training process, forming a teacher group; this teacher group then transmits knowledge to the student group through three teaching methods (intensive teaching, private teaching, inter-group teaching).
Ord2Seq: Regarding Ordinal Regression as Label Sequence Prediction
Jinhong Wang (Zhejiang University), Jian Wu (Zhejiang University)
ClassificationTransformerImage
🎯 What it does: Transform ordinal labels into binary sequences, using a Transformer encoder-decoder structure for recursive binary classification, ultimately obtaining the original ordinal predictions.
Order-preserving Consistency Regularization for Domain Adaptation and Generalization
Mengmeng Jing (University of Electronic Science and Technology of China), Cees G. M. Snoek (University of Amsterdam)
Domain AdaptationImage
🎯 What it does: A method of Order Consistency Regularization (OCR) is proposed, which reduces the model's sensitivity to domain-specific attributes by maximizing the entropy of the difference between the representations of the original and augmented images.
Order-Prompted Tag Sequence Generation for Video Tagging
Zongyang Ma (Chinese Academy of Sciences), Weiming Hu (Chinese Academy of Sciences)
ClassificationGenerationTransformerContrastive LearningVideoText
🎯 What it does: The OP-TSG model is proposed, treating video tags as a sequence of words generated in order, and decoupling the tag order through sample-dependent sequential prompts, thereby enhancing the modeling of tag dependencies and the ability to generate new tags.
Ordered Atomic Activity for Fine-grained Interactive Traffic Scenario Understanding
Nakul Agarwal (Honda Research Institute), Yi-Ting Chen (National Yang Ming Chiao Tung University)
RecognitionRetrievalGraph Neural NetworkVideo
🎯 What it does: This study proposes the Ordered Atomic Activity representation method, which splits interactive scenes into ordered atomic activities based on road topology and constructs the OATS dataset.
Ordinal Label Distribution Learning
Changsong Wen (Nankai University), Jufeng Yang (Nankai University)
ClassificationRecognitionConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: This paper proposes a new label distribution learning paradigm—Ordered Label Distribution Learning (OLDL), and designs three order-sensitive learning objectives and evaluation metrics (CAD, QFD, CJS) specifically for tasks where labels have a natural order.
OrthoPlanes: A Novel Representation for Better 3D-Awareness of GANs
Honglin He (Tsinghua University), Wayne Wu (Shanghai AI Laboratory)
GenerationData SynthesisPose EstimationGenerative Adversarial NetworkImage
🎯 What it does: A new hybrid explicit-implicit 3D representation called OrthoPlanes is proposed, which improves the 3D perception of 2D GANs and is capable of generating geometrically detailed and viewpoint-consistent images from multiple perspectives.
Out-of-Distribution Detection for Monocular Depth Estimation
Julia Hornauer (Ulm University), Vasileios Belagiannis (Friedrich-Alexander-Universität Erlangen-Nürnberg)
Depth EstimationAnomaly DetectionConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a method that builds on a trained monocular depth estimation model by adding a post-training image reconstruction decoder, using reconstruction error to detect out-of-distribution (OOD) inputs, thereby enhancing the safety of depth estimation.
Out-of-Domain GAN Inversion via Invertibility Decomposition for Photo-Realistic Human Face Manipulation
Xin Yang (Hong Kong University of Science and Technology), Yingcong Chen (Hong Kong University of Science and Technology)
Image TranslationGenerationGenerative Adversarial NetworkOptical FlowImage
🎯 What it does: This paper proposes a facial image GAN inversion and attribute editing method achieved through Invertibility Decomposition, which can partition the input image into an invertible inner domain (ID) and a non-invertible outer domain (OOD), and directly concatenate the OOD region with the input image to obtain high-fidelity, ghosting-free synthesis results.
Overcoming Forgetting Catastrophe in Quantization-Aware Training
Ting-An Chen (National Taiwan University), Ming-Syan Chen (National Taiwan University)
ClassificationRecognitionImage
🎯 What it does: This paper proposes a lifelong quantization method called LifeQuant, which can continuously perform quantization training on streaming data and effectively avoid catastrophic forgetting.
Overwriting Pretrained Bias with Finetuning Data
Angelina Wang (Princeton University), Olga Russakovsky (Princeton University)
ClassificationRepresentation LearningConvolutional Neural NetworkSupervised Fine-TuningContrastive LearningImage
🎯 What it does: This study investigates how biases (such as spurious correlations and underrepresentation of minority groups) in pre-trained models are transferred after fine-tuning, and explores the feasibility of mitigating these biases by adjusting the distribution of fine-tuning data.
OxfordTVG-HIC: Can Machine Make Humorous Captions from Images?
Runjia Li (University of Oxford), Philip Torr (University of Oxford)
GenerationTransformerLarge Language ModelVision Language ModelImageText
🎯 What it does: A humorous image caption dataset, OxfordTVG-HIC, consisting of 2.89 million image-text pairs, was constructed, and a humor generation model was trained based on this dataset.
P1AC: Revisiting Absolute Pose From a Single Affine Correspondence
Jonathan Ventura (California Polytechnic State University), Dániel Baráth (ETH Zurich)
Pose EstimationRetrievalSimultaneous Localization and MappingImageBenchmark
🎯 What it does: This paper proposes a new method for solving the absolute pose of a camera using a single affine correspondence (P1AC), which enables camera localization based solely on one 3D point, its surface normal vector, and the pose of a reference image.
P2C: Self-Supervised Point Cloud Completion from Single Partial Clouds
Ruikai Cui (Australian National University), Nick Barnes (Australian National University)
RestorationGenerationPoint Cloud
🎯 What it does: A completely self-supervised point cloud completion framework P2C is proposed, which can train and complete the full shape using only a single incomplete point cloud of each object.
PADCLIP: Pseudo-labeling with Adaptive Debiasing in CLIP for Unsupervised Domain Adaptation
Zhengfeng Lai (University of California), Chen-Nee Chuah (University of California)
Domain AdaptationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes a framework PADCLIP that uses CLIP in unsupervised domain adaptation, combining pseudo-labeling and adaptive debiasing methods.
PADDLES: Phase-Amplitude Spectrum Disentangled Early Stopping for Learning with Noisy Labels
Huaxi Huang (Data61 CSIRO), Tongliang Liu (University of Sydney)
ClassificationData-Centric LearningTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes the PADDLES method, which utilizes discrete Fourier transform to decompose intermediate features of the network into amplitude spectrum and phase spectrum, and sets different early stopping points for both to enhance robustness in noisy label environments.
Pairwise Similarity Learning is SimPLE
Yandong Wen (Max Planck Institute for Intelligent Systems), Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)
ClassificationRecognitionRetrievalContrastive LearningImage
🎯 What it does: This paper proposes a new method called SimPLE for learning pairwise similarity (PSL), which aims to assign higher similarity scores to positive sample pairs (same label) and lower scores to negative sample pairs (different labels).
PanFlowNet: A Flow-Based Deep Network for Pan-Sharpening
Gang Yang (University of Science and Technology of China), Deyu Meng (Nanyang Technological University)
RestorationGenerationData SynthesisFlow-based ModelImage
🎯 What it does: A flow model-based PanFlowNet is proposed for the fusion of multispectral and panchromatic images, capable of generating diverse high-resolution multispectral images.
Panoramas from Photons
Sacha Jungerman (University of Wisconsin-Madison), Mohit Gupta (University of Wisconsin-Madison)
RestorationGenerationSuper ResolutionOptical FlowImageVideo
🎯 What it does: Using binary frames generated by a single-photon camera, an iterative resampling and aggregation method is proposed to estimate global motion at extremely high speeds, thereby generating high-quality panoramic images, super-resolution images, and high dynamic range images.
Parallax-Tolerant Unsupervised Deep Image Stitching
Lang Nie (Beijing Jiaotong University), Yao Zhao (Beijing Jiaotong University)
Image TranslationImage HarmonizationOptimizationConvolutional Neural NetworkImage
🎯 What it does: A novel unsupervised deep image stitching method is proposed, capable of handling both large parallax and low-texture scenes, avoiding the limitations of traditional handcrafted geometric features.
Parallel Attention Interaction Network for Few-Shot Skeleton-Based Action Recognition
Xingyu Liu (Xi'an Jiaotong University), Gang Hua (Wormpex AI Research)
RecognitionPose EstimationGraph Neural NetworkVideo
🎯 What it does: A parallel attention interaction network (PAINet) is proposed and implemented for skeleton action recognition under very few samples, enhancing the discriminability of action features through spatial cross-attention, spatial self-attention, and time series matching.
Parameterized Cost Volume for Stereo Matching
Jiaxi Zeng (Beijing Institute of Technology), Yunde Jia (Shenzhen MSU-BIT University)
Depth EstimationAutonomous DrivingOptimizationComputational EfficiencyRecurrent Neural NetworkGaussian SplattingImage
🎯 What it does: This paper proposes a cost volume parameterized by multiple Gaussian distributions to encode the entire disparity space, achieving rapid iterative convergence through JS-divergence optimization, significantly enhancing stereo matching speed.
Parametric Classification for Generalized Category Discovery: A Baseline Study
Xin Wen (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)
ClassificationKnowledge DistillationRepresentation LearningContrastive LearningImageBenchmark
🎯 What it does: This paper investigates the reasons for the failure of parametric classifiers in the task of Generalized Category Discovery (GCD) and proposes a simple and efficient baseline (SimGCD) based on self-distillation and entropy regularization. By improving pseudo-labels, joint training, and post-backbone features, it significantly enhances the ability to discover new categories.
Parametric Depth Based Feature Representation Learning for Object Detection and Segmentation in Bird's-Eye View
Jiayu Yang (Australian National University), Jose M. Alvarez (NVIDIA)
Object DetectionSegmentationAutonomous DrivingRepresentation LearningMultimodality
🎯 What it does: Using parameterized depth distribution to upsample multi-view image features to BEV space, jointly achieving 3D object detection and BEV semantic segmentation, and generating visibility maps to suppress hallucinations.
Parametric Information Maximization for Generalized Category Discovery
Florent Chiaroni (Thales Digital Solutions), Ismail Ben Ayed (Thales Digital Solutions)
ClassificationOptimizationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: Proposes the Parametric Information Maximization (PIM) model, targeting the Generalized Category Discovery (GCD) task, utilizing dual-layer optimization and supervised constraints to achieve mutual information maximization between features and labels.
ParCNetV2: Oversized Kernel with Enhanced Attention
Ruihan Xu (Peking University), Xiaoyu Wang (Peking University)
Object DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: A pure convolutional network ParCNetV2 is proposed, which efficiently models global context and attention mechanisms by using oversized convolution and bifurcate gate units.
PARF: Primitive-Aware Radiance Fusion for Indoor Scene Novel View Synthesis
Haiyang Ying (Tsinghua University), Lu Fang (Tsinghua University)
GenerationData SynthesisNeural Radiance FieldPoint Cloud
🎯 What it does: A primitive-aware radiance fusion framework named PARF is proposed for fast radiance field reconstruction and editing under sparse RGB-D input in indoor scenes.
PARIS: Part-level Reconstruction and Motion Analysis for Articulated Objects
Jiayi Liu (Simon Fraser University), Manolis Savva (Simon Fraser University)
Object DetectionPose EstimationNeural Radiance FieldImagePoint Cloud
🎯 What it does: Using two sets of multi-view RGB images, we self-supervisedly reconstruct the shape and appearance of joint object parts and estimate their motion parameters without 3D/motion annotations.
Part-Aware Transformer for Generalizable Person Re-identification
Hao Ni (University of Electronic Science and Technology of China), Jingkuan Song (Shenzhen Institute for Advanced Study)
RecognitionRetrievalDomain AdaptationKnowledge DistillationTransformerContrastive LearningImage
🎯 What it does: This paper proposes a pure Transformer framework called Part-Aware Transformer, designed for domain generalization in person re-identification, incorporating two proxy tasks: Cross-ID Similarity Learning and Part-Guided Self-Distillation to enhance cross-domain generalization capability.
Partition Speeds Up Learning Implicit Neural Representations Based on Exponential-Increase Hypothesis
Ke Liu (Zhejiang University), Bo Han (Hong Kong Baptist University)
Super ResolutionOptimizationMeta LearningImageComputed Tomography
🎯 What it does: The paper proposes and validates the exponential growth hypothesis, which states that the convergence speed of network fitting for fault images increases exponentially with the number of boundaries. It reduces this to linear through partitioning, thereby accelerating INR learning.
Partition-And-Debias: Agnostic Biases Mitigation via a Mixture of Biases-Specific Experts
Jiaxuan Li (University of Tokyo), Hideki Nakayama (University of Tokyo)
ClassificationMixture of ExpertsImage
🎯 What it does: Proposes an agnostic biases scenario and achieves automatic decomposition and elimination of various unknown biases through the Partition-and-Debias (PnD) method, thereby enabling unbiased image classification.
PARTNER: Level up the Polar Representation for LiDAR 3D Object Detection
Ming Nie (Fudan University), Li Zhang (Fudan University)
Object DetectionAutonomous DrivingTransformerPoint Cloud
🎯 What it does: This paper proposes PARTNER, an end-to-end framework for LiDAR 3D object detection in polar coordinates, which addresses the feature distortion problem in polar coordinate representation through Global Feature Re-alignment (GRR) and Geometric Adaptation (GA) modules.
Passive Ultra-Wideband Single-Photon Imaging
Mian Wei (University of Toronto), Kiriakos N. Kutulakos (University of Toronto)
Image TranslationRestorationObject DetectionVideoStochastic Differential Equation
🎯 What it does: This paper proposes a non-synchronous, passive single-photon imaging framework that can image dynamic scenes on extreme time scales from seconds to picoseconds. By constructing the 'flux probing' theory, it utilizes the absolute timestamp stream from single-photon avalanche diodes (SPADs) to reconstruct the time-varying flux function of pixels, and implements a corresponding frequency domain reconstruction algorithm that supports broadband flux estimation from DC to approximately 31 GHz.
PASTA: Proportional Amplitude Spectrum Training Augmentation for Syn-to-Real Domain Generalization
Prithvijit Chattopadhyay (Georgia Institute of Technology), Judy Hoffman (Georgia Institute of Technology)
Object DetectionSegmentationDomain AdaptationImage
🎯 What it does: In the domain transfer task from synthetic data to real data, a training enhancement method based on frequency domain amplitude spectrum ratio perturbation, called PASTA, is proposed;
PatchCT: Aligning Patch Set and Label Set with Conditional Transport for Multi-Label Image Classification
Miaoge Li (Xidian University), Mingyuan Zhou (University of Texas at Austin)
ClassificationTransformerPrompt EngineeringImageText
🎯 What it does: Transform the multi-label image classification problem into a conditional transport (CT) problem of two discrete distributions: visual patches and text labels, and align the two modalities by minimizing the bidirectional CT distance.