ICCV 2023 Papers — Page 9
IEEE/CVF International Conference on Computer Vision · 2156 papers
GlobalMapper: Arbitrary-Shaped Urban Layout Generation
Liu He (Purdue University), Daniel Aliaga (Purdue University)
GenerationGraph Neural NetworkAuto EncoderGraph
🎯 What it does: A controllable urban layout generation framework based on graph attention networks has been developed, capable of generating diverse building layouts under any road network conditions and supporting prior condition generation.
Gloss-Free Sign Language Translation: Improving from Visual-Language Pretraining
Benjia Zhou (MUST), Du Zhang (MUST)
RecognitionGenerationTransformerVision Language ModelContrastive LearningVideoText
🎯 What it does: This paper proposes a gloss-free supervised sign language translation framework GFSLT-VLP, which utilizes visual-language pre-training (VLP) and masked self-supervised learning to enhance the language-guided representation of the visual encoder, thereby achieving end-to-end translation from sign language to text.
GlowGAN: Unsupervised Learning of HDR Images from LDR Images in the Wild
Chao Wang (MPI Informatik), Thomas Leimkühler (Nanyang Technological University)
RestorationGenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: An unsupervised learning method for generating high dynamic range (HDR) images from low dynamic range (LDR) images, called GlowGAN, is proposed.
GlueGen: Plug and Play Multi-modal Encoders for X-to-image Generation
Can Qin (Northeastern University), Ran Xu (Salesforce AI Research)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkImageMultimodalityAudio
🎯 What it does: This paper proposes the GlueGen framework, which utilizes GlueNet to achieve seamless alignment of different conditional encoders (such as multilingual and audio) with existing stable diffusion models, enabling X-to-image generation.
GlueStick: Robust Image Matching by Sticking Points and Lines Together
Rémi Pautrat (ETH Zurich), Viktor Larsson (Qualcomm)
Object DetectionPose EstimationGraph Neural NetworkImage
🎯 What it does: This paper proposes GlueStick, which utilizes Graph Neural Networks (GNN) to simultaneously match keypoints and line segments within a single network, and constructs a wireframe structure based on the connectivity of line segment endpoints.
GO-SLAM: Global Optimization for Consistent 3D Instant Reconstruction
Youmin Zhang (University of Bologna), Matteo Poggi (University of Bologna)
Pose EstimationDepth EstimationOptimizationRobotic IntelligenceSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: A real-time, deep learning-driven global consistent 3D visual SLAM system called GO-SLAM is proposed, capable of precise localization and high-quality scene reconstruction under monocular, stereo, or RGB-D sensors.
Going Beyond Nouns With Vision & Language Models Using Synthetic Data
Paola Cascante-Bonilla (Rice University), Leonid Karlinsky (IBM Research)
Data SynthesisDomain AdaptationTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: By generating millions of synthetic image-text pairs (SyViC dataset) and combining techniques such as parameter-efficient fine-tuning (LoRA), domain-adaptive style transfer, and long text chunking, we fine-tune existing large-scale vision-language models (such as CLIP, CyCLIP) to enhance their understanding of visual language concepts beyond nouns (attributes, relationships, states) and their ability for compositional reasoning.
Going Denser with Open-Vocabulary Part Segmentation
Peize Sun (University of Hong Kong), Zhicheng Yan (University of Hong Kong)
Object DetectionSegmentationTransformerContrastive LearningImage
🎯 What it does: A detector capable of performing object and part segmentation at open vocabulary and multi-level fine granularity is proposed;
GPA-3D: Geometry-aware Prototype Alignment for Unsupervised Domain Adaptive 3D Object Detection from Point Clouds
Ziyu Li (Southeast University), Wankou Yang (Southeast University)
Object DetectionDomain AdaptationAutonomous DrivingContrastive LearningPoint Cloud
🎯 What it does: This paper proposes an unsupervised domain adaptation framework GPA-3D, which utilizes the geometric structure of point clouds to align prototypes of BEV features, thereby reducing cross-domain differences and improving the performance of LiDAR-based 3D detection.
GPFL: Simultaneously Learning Global and Personalized Feature Information for Personalized Federated Learning
Jianqing Zhang (Shanghai Jiao Tong University), Haibing Guan (Shanghai Jiao Tong University)
Federated LearningSafty and PrivacyConvolutional Neural NetworkContrastive LearningImageText
🎯 What it does: A framework called GPFL is proposed for federated learning that simultaneously learns global and personalized feature information, achieving personalized models while ensuring collaborative learning.
GPGait: Generalized Pose-based Gait Recognition
Yang Fu (Beijing Normal University), Yongzhen Huang (Beijing Normal University)
RecognitionPose EstimationDomain AdaptationGraph Neural NetworkImageVideo
🎯 What it does: This paper proposes the GPGait framework, which first unifies and enhances pose representation using Human Orientation Transformation (HOT) and Human Orientation Descriptor (HOD), and then learns local-global relationships through a Part-Aware Graph Convolutional Network (PAGCN) to achieve cross-domain gait recognition.
Gradient-based Sampling for Class Imbalanced Semi-supervised Object Detection
Jiaming Li (Sun Yat-sen University), Guanbin Li (Sun Yat-sen University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: An experimental setup for class-imbalanced semi-supervised object detection (CI-SSOD) is proposed, along with a gradient-based sampling framework to address two types of confirmation bias.
Gradient-Regulated Meta-Prompt Learning for Generalizable Vision-Language Models
Juncheng Li (Zhejiang University), Yueting Zhuang (Zhejiang University)
Domain AdaptationMeta LearningTransformerPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: This study investigates the initialization sensitivity and overfitting issues of prompt tuning in few-shot scenarios, proposing the Gradient-Regulated Meta-Prompt learning (GRAM) framework. It jointly learns the initialization of soft prompts and gradient regulation functions, achieving meta-learning using only unlabeled image-text pairs, thereby enhancing the model's adaptability and generalization ability in new tasks and domains.
Gram-based Attentive Neural Ordinary Differential Equations Network for Video Nystagmography Classification
Xihe Qiu (Shanghai University of Engineering Science), Huawei Li (Shanghai University of Engineering Science)
ClassificationVideoTime SeriesOrdinary Differential Equation
🎯 What it does: This paper proposes an end-to-end BPPV diagnostic framework TC-BPPV, which transforms video nystagmography (VNG) into eye movement trajectories and utilizes a Gram-AODE network for classification, ultimately achieving automatic diagnosis of BPPV types.
GRAM-HD: 3D-Consistent Image Generation at High Resolution with Generative Radiance Manifolds
Jianfeng Xiang (Tsinghua University), Xin Tong (Microsoft Research Asia)
GenerationData SynthesisSuper ResolutionConvolutional Neural NetworkNeural Radiance FieldGenerative Adversarial NetworkImage
🎯 What it does: A GAN model called GRAM-HD is proposed, which can generate high-resolution (1024×1024) images while maintaining strict 3D consistency.
Gramian Attention Heads are Strong yet Efficient Vision Learners
Jongbin Ryu (Ajou University), Jongwoo Lim (Seoul National University)
ClassificationSegmentationConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: In visual classification tasks, a multi-head lightweight classifier is introduced, and Gramian attention is used to enhance class labels, improving the model's expressive capability.
Graph Matching with Bi-level Noisy Correspondence
Yijie Lin (Sichuan University), Xi Peng (Sichuan University)
OptimizationKnowledge DistillationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: The Bi-Level Noisy Correspondence (BNC) problem in graph matching is proposed, and the COMMON method is introduced to achieve robust matching through contrastive learning and momentum distillation.
GraphAlign: Enhancing Accurate Feature Alignment by Graph matching for Multi-Modal 3D Object Detection
Ziying Song (Beijing Jiaotong University), Caiyan Jia (Beijing Jiaotong University)
Object DetectionAutonomous DrivingConvolutional Neural NetworkGraph Neural NetworkSupervised Fine-TuningMultimodalityPoint Cloud
🎯 What it does: This paper proposes GraphAlign, which utilizes graph matching to achieve precise alignment of LiDAR and camera features, and injects the aligned features into a 3D detection network.
GraphEcho: Graph-Driven Unsupervised Domain Adaptation for Echocardiogram Video Segmentation
Jiewen Yang (Hong Kong University of Science and Technology), Xiaomeng Li (Hong Kong University of Science and Technology)
SegmentationDomain AdaptationGraph Neural NetworkContrastive LearningVideoBiomedical DataUltrasound
🎯 What it does: A graph-based unsupervised domain adaptation method called GraphEcho is proposed for structural segmentation of cardiac ultrasound videos.
Graphics2RAW: Mapping Computer Graphics Images to Sensor RAW Images
Donghwan Seo (Samsung Electronics), Michael S. Brown (Samsung AI Center Toronto)
Image TranslationRestorationImage
🎯 What it does: Using a small number of RAW-DNG files from the target sensor, by sampling the lighting and reverse inferring the ISP process, the computer-generated sRGB images are mapped to realistic sensor RAW images, supporting multiple sensors and various lighting conditions.
GridMM: Grid Memory Map for Vision-and-Language Navigation
Zihan Wang (Chinese Academy of Sciences), Shuqiang Jiang (Chinese Academy of Sciences)
TransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: A dynamic growth grid memory map (GridMM) is proposed to structurally record historical environments in visual and language navigation tasks, and an instruction-related aggregation method is designed to capture fine-grained visual cues.
GridPull: Towards Scalability in Learning Implicit Representations from 3D Point Clouds
Chao Chen (Tsinghua University), Zhizhong Han (Wayne State University)
OptimizationRepresentation LearningPoint Cloud
🎯 What it does: The GridPull method is proposed, which achieves efficient surface reconstruction of large-scale point clouds by directly optimizing the distance field on a discrete grid, avoiding the use of neural networks.
Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation
Yibo Cui (Defense Innovation Institute), Erwei Yin (Defense Innovation Institute)
TransformerVision Language ModelMultimodality
🎯 What it does: By constructing a high-quality entity-landmark alignment dataset GEL-R2R, and based on this, conducting three entity-landmark level adaptive pre-training tasks (entity phrase prediction, landmark box prediction, semantic alignment) for the VLN pre-training model, the fine-grained cross-modal alignment ability of visual and language navigation is enhanced.
Grounded Image Text Matching with Mismatched Relation Reasoning
Yu Wu (ShanghaiTech University), Xuming He (ShanghaiTech University)
Graph Neural NetworkTransformerVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A new visual-language joint task is proposed - GITM-MR (Grounded Image Text Matching with Mismatched Relation), along with a corresponding evaluation benchmark;
Grounding 3D Object Affordance from 2D Interactions in Images
Yuhang Yang (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
RecognitionObject DetectionRobotic IntelligenceConvolutional Neural NetworkTransformerImageMultimodalityPoint Cloud
🎯 What it does: Proposes to locate interactive areas in 3D point clouds from 2D interactive images, addressing the problem of operability recognition for 3D objects.
Group DETR: Fast DETR Training with Group-Wise One-to-Many Assignment
Qiang Chen (Baidu), Jingdong Wang (Baidu)
Object DetectionSegmentationAutonomous DrivingTransformerImagePoint Cloud
🎯 What it does: The Group DETR method is proposed, which accelerates the training of DETR and improves performance by introducing multiple groups of object queries and one-to-many matching within groups.
Group Pose: A Simple Baseline for End-to-End Multi-Person Pose Estimation
Huan Liu (Beijing Jiaotong University), Jingdong Wang (Baidu)
Pose EstimationTransformerImage
🎯 What it does: A simple end-to-end multi-person pose estimation framework called Group Pose is proposed.
GrowCLIP: Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training
Xinchi Deng (Sun Yat-sen University), Xiaodan Liang
RetrievalData-Centric LearningNeural Architecture SearchTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A contrastive language-image pre-training model called GrowCLIP is designed for online learning scenarios, capable of automatically expanding the network scale and selecting the optimal architecture based on the continuously growing data.
Growing a Brain with Sparsity-Inducing Generation for Continual Learning
Hyundong Jin (Chung Ang University), Eunwoo Kim (Chung Ang University)
ClassificationRecognitionTransformerImageVideo
🎯 What it does: This paper proposes GrowBrain, which uses a hypernetwork to dynamically evolve the parameters of old tasks in continual learning, avoiding the problem of knowledge stagnation caused by fixed parameters.
Guided Motion Diffusion for Controllable Human Motion Synthesis
Korrawe Karunratanakul (ETH Zurich), Siyu Tang (ETH Zurich)
GenerationData SynthesisPose EstimationDiffusion modelVideoText
🎯 What it does: This paper proposes Guided Motion Diffusion (GMD), a controllable human motion synthesis method based on diffusion models, which can introduce spatial constraints such as trajectories, keyframes, and obstacles based on text prompts.
Guiding Image Captioning Models Toward More Specific Captions
Simon Kornblith (Google DeepMind), Thao Nguyen (University of Washington)
GenerationRetrievalTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: This study introduces classifier-free guidance (CFG) and language model guidance into the image captioning model to make the generated captions more specific, and explores the impact of different guidance scales on generation quality.
Guiding Local Feature Matching with Surface Curvature
Shuzhe Wang (Aalto University), Daniel Barath (ETH Zurich)
Pose EstimationDepth EstimationTransformerSupervised Fine-TuningImagePoint Cloud
🎯 What it does: A feature matching guidance method based on surface curvature, CSE, is proposed, which utilizes the three-dimensional curvature information predicted by monocular depth to enhance the accuracy of local feature matching.
H3WB: Human3.6M 3D WholeBody Dataset and Benchmark
Yue Zhu (Ecole des Ponts), David Picard (Ecole des Ponts)
Pose EstimationTransformerDiffusion modelAuto EncoderImageBenchmark
🎯 What it does: The Human3.6M 3D WholeBody (H3WB) dataset is proposed, and based on this, three benchmark tasks for 3D full-body pose estimation are defined; multi-stage baseline results are also provided.
HairCLIPv2: Unifying Hair Editing via Proxy Feature Blending
Tianyi Wei (University of Science and Technology of China), Nenghai Yu (University of Science and Technology of China)
Image TranslationGenerationGenerative Adversarial NetworkImage
🎯 What it does: A unified hair editing framework called HairCLIPv2 is proposed, which supports global or local hairstyle and hair color editing using text, reference images, sketches, masks, RGB, and their combinations.
HairNeRF: Geometry-Aware Image Synthesis for Hairstyle Transfer
Seunggyu Chang (NAVER Cloud), Hayeon Kim (UNIST)
Image TranslationGenerationData SynthesisNeural Radiance FieldGenerative Adversarial NetworkImage
🎯 What it does: We propose HairNeRF, a geometry-aware neural rendering framework for hairstyle transfer across viewpoints and identities, capable of achieving high-quality image synthesis while preserving the geometric structure of the head.
HAL3D: Hierarchical Active Learning for Fine-Grained 3D Part Labeling
Fenggen Yu (Amazon), Hao Zhang (Simon Fraser University)
Object DetectionSegmentationGraph Neural NetworkPoint Cloud
🎯 What it does: A human-computer interactive active learning framework HAL3D is proposed for precise labeling of fine-grained components of 3D models, iteratively validating or modifying the predictions of deep networks and using feedback for model fine-tuning.
Hallucination Improves the Performance of Unsupervised Visual Representation Learning
Jing Wu (University of Illinois Urbana Champaign), Naira Hovakimyan (University of Illinois Urbana Champaign)
Object DetectionSegmentationRepresentation LearningContrastive LearningImage
🎯 What it does: A module named Hallucinator is proposed, which can generate additional hard positive samples in the feature space through differential inference and nonlinear transformations, thereby enhancing the diversity of positive samples in contrastive learning.
HaMuCo: Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning
Xiaozheng Zheng (Beijing University of Posts and Telecommunications), Jingyu Wang (Beijing University of Posts and Telecommunications)
Pose EstimationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A self-supervised learning framework HaMuCo is proposed, which trains a single-view 3D hand pose estimator using multi-view pseudo 2D labels and achieves collaborative learning through a cross-view interaction network.
HandR2N2: Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network
Wencan Cheng (Sungkyunkwan University), Jong Hwan Ko (Sungkyunkwan University)
Pose EstimationRecurrent Neural NetworkGraph Neural NetworkPoint Cloud
🎯 What it does: Iteratively estimating 3D hand poses from a single point cloud using residual recursive units
Handwritten and Printed Text Segmentation: A Signature Case Study
Sina Gholamian (Thomson Reuters AI Labs), Ali Vahdat (Thomson Reuters AI Labs)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: Four types of pixel-level segmentation methods (handwritten, printed, background, overlapping) are proposed, and a high-quality handwritten and printed overlapping document dataset, SignaTR6K, is provided.
Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient
Zhengzhi Lu (Xi'an Jiaotong University), Hubert P. H. Shum (Durham University)
RecognitionAdversarial AttackContrastive LearningTime SeriesSequential
🎯 What it does: A method for adversarial attacks on skeleton action recognition models is proposed in a hard black-box attack environment without models, training data, or labels.
Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis
Qiucheng Wu (University of California Santa Barbara), Shiyu Chang (Adobe Research)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: Achieving high-fidelity text-to-image synthesis by explicitly controlling spatio-temporal cross-attention in diffusion models.
Harvard Glaucoma Detection and Progression: A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning
Yan Luo (Harvard University), Mengyu Wang (Harvard University)
ClassificationSegmentationAnomaly DetectionConvolutional Neural NetworkTransformerReinforcement LearningImageMultimodality
🎯 What it does: This paper proposes a Pseudo-Supervisor based on reinforcement learning, which generates pseudo-labels using unlabeled OCT data through a policy network to enhance the performance of glaucoma detection and progression prediction for retinal nerve fiber layer thickness (RNFLT) images.
Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time
Cheng-Hung Chan (National Tsing Hua University), Hwann-Tzong Chen (National Tsing Hua University)
GenerationComputational EfficiencyOptical FlowVideo
🎯 What it does: A layer-based neural video decomposition method is proposed, utilizing hash grid encoding and multiplicative residual estimation to achieve fast training (about 40 minutes) and real-time rendering (71fps) for high-resolution videos, while supporting editing of lighting changes.
HDG-ODE: A Hierarchical Continuous-Time Model for Human Pose Forecasting
Yucheng Xing (Stony Brook University), Xin Wang (Stony Brook University)
Pose EstimationGraph Neural NetworkTime SeriesSequentialOrdinary Differential Equation
🎯 What it does: This study proposes a continuous-time model based on hierarchical dynamic graph ODE for predicting future 3D human poses from multi-person 2D skeleton sequences.
Helping Hands: An Object-Aware Ego-Centric Video Recognition Model
Chuhan Zhang (VGG University of Oxford), Andrew Zisserman (VGG University of Oxford)
RecognitionObject DetectionRetrievalTransformerContrastive LearningVideoText
🎯 What it does: This paper proposes an object-aware autoregressive decoder that enhances the spatial-temporal representation of first-person videos during the training phase using weak supervision from hand and object localization and semantic labels; during inference, tasks such as video-text retrieval, action classification, and long-term video understanding can be completed using only RGB frames.
Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking
Rui Li (Beijing Jiaotong University), Zhu Teng (Beijing Jiaotong University)
Object TrackingReinforcement LearningVideo
🎯 What it does: A heterogeneous diversity-driven active learning framework based on Markov Decision Process (HD-AMOT) is proposed for efficiently selecting valuable frames for annotation in single-stage multi-object tracking.
Heterogeneous Forgetting Compensation for Class-Incremental Learning
Jiahua Dong (State Key Laboratory of Robotics, Shenyang Institute of Automation, Chinese Academy of Sciences), Gan Sun (State Key Laboratory of Robotics, Shenyang Institute of Automation, Chinese Academy of Sciences)
ClassificationKnowledge DistillationTransformerImage
🎯 What it does: The Heterogeneous Forgetting Compensation (HFC) model is proposed in class-incremental learning to address the different forgetting rates of old classes at both the representation and gradient levels.
Hidden Biases of End-to-End Driving Models
Bernhard Jaeger (University of Tuebingen), Andreas Geiger (University of Tuebingen)
Autonomous DrivingTransformerPoint Cloud
🎯 What it does: Analyze the hidden biases of existing end-to-end driving models and propose an improved TransFuser++ system based on this analysis, significantly enhancing CARLA evaluation metrics.
Hiding Visual Information via Obfuscating Adversarial Perturbations
Zhigang Su (Xidian University), Xinbo Gao (Chongqing University of Posts and Telecommunications)
ClassificationRecognitionAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A visual information hiding method based on Type-I adversarial attacks (AVIH) is designed to completely obscure the visual content of images without altering the service model, while maintaining the original functionalities (such as face recognition and classification).
Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection
Xin Feng (Harbin Institute of Technology), Wenjie Pei (Harbin Institute of Technology)
RestorationAnomaly DetectionTransformerGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: A hierarchical contrastive learning framework is proposed for blind detection of damaged regions in images and non-destructive image restoration.
Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models
Huaijin Pi (Zhejiang University), Hujun Bao (Zhejiang University)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: For a given object, first generate the target pose, then predict a series of milestones, and finally synthesize complete 3D human-machine interaction movements between the milestones using a diffusion model, achieving long-term, diverse human action generation.
Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation
Zongyi Xu (Chongqing University of Posts and Telecommunications), Xinbo Gao (Chongqing University of Posts and Telecommunications)
SegmentationKnowledge DistillationConvolutional Neural NetworkPoint Cloud
🎯 What it does: This paper proposes a hierarchical point-based active learning framework for semi-supervised 3D point cloud semantic segmentation, which can significantly improve segmentation performance with very few labeled points.
Hierarchical Prior Mining for Non-local Multi-View Stereo
Chunlin Ren (Northwestern Polytechnical University), Jiaqi Yang (Northwestern Polytechnical University)
RestorationDepth EstimationSimultaneous Localization and MappingImageBenchmark
🎯 What it does: A multi-view stereo reconstruction method based on hierarchical prior mining, HPM-MVS, is proposed, which integrates non-local expandable sampling patterns (NESP), KNN-based plane prior construction, and a multi-scale hierarchical prior mining framework.
Hierarchical Spatio-Temporal Representation Learning for Gait Recognition
Lei Wang (Hebei Agricultural University), Bincheng Wang (Hebei Agricultural University)
RecognitionRepresentation LearningConvolutional Neural NetworkVideo
🎯 What it does: This paper proposes a Hierarchical Spatio-Temporal Representation Learning Framework (HSTL), which extracts human motion features layer by layer from coarse to fine through an Adaptive Region Motion Extractor (ARME), Adaptive Spatio-Temporal Pooling (ASTP), and Frame-level Temporal Aggregation (FTA) modules.
Hierarchical Visual Categories Modeling: A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection
Jinglun Li (Fudan University), Wenqiang Zhang (Fudan University)
Anomaly DetectionKnowledge DistillationRepresentation LearningImage
🎯 What it does: Proposes a Hierarchical Visual Category Modeling (HVCM) framework that jointly trains visual representations and multiple sets of Gaussian mixture models to achieve OOD detection without external anomalous samples.
Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning
Hanjae Kim (Yonsei University), Kwanghoon Sohn (Korea Institute of Science and Technology)
ClassificationRecognitionTransformerMixture of ExpertsImage
🎯 What it does: A Composition Transformer (CoT) framework based on Transformer is proposed, which combines object experts and attribute experts to achieve hierarchical separation and fusion of attribute and object features through object-guided attention.
Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition
Jungho Lee (Yonsei University), Sangyoun Lee (AIonFlow Research)
ClassificationRecognitionGraph Neural NetworkVideoGraph
🎯 What it does: A hierarchical split graph convolutional network (HD-GCN) is proposed for skeletal action recognition, constructing a hierarchical split graph (HD-Graph) and combining it with an attention aggregation module (A-HA), followed by a six-stream non-motion flow ensemble to enhance performance.
HiFace: High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details
Zenghao Chai (National University of Singapore), Jiang Bian (Microsoft Research Asia)
GenerationData SynthesisKnowledge DistillationImage
🎯 What it does: A complete system for reconstructing high-fidelity 3D faces from a single image, enabling the animation of both static and dynamic details.
High Quality Entity Segmentation
Lu Qi (Adobe Research), Ming-Hsuan Yang (Adobe Research)
Object DetectionSegmentationTransformerImageBenchmark
🎯 What it does: A high-quality, cross-domain, large-scale high-resolution entity segmentation dataset called EntitySeg has been constructed, and a Transformer structure named CropFormer has been proposed, which significantly improves entity segmentation accuracy by integrating full image and local crop information.
High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net
Zinuo Li (University of Macau), Xiaodong Cun (University of Macau)
RestorationTransformerGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a document shadow removal network based on frequency domain decomposition, called FSENet, and constructs a dataset of 7k pairs of high-resolution shadowed documents, named SD7K.
HiLo: Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation
Zijian Zhou (King's College London), Holger Caesar (Delft University of Technology)
Object DetectionGenerationTransformerGraph
🎯 What it does: This paper addresses the long-tail relationship and semantic overlap issues in scene graph generation by proposing the HiLo framework.
HiTeA: Hierarchical Temporal-Aware Video-Language Pre-training
Qinghao Ye (DAMO Academy Alibaba Group), Fei Huang (DAMO Academy Alibaba Group)
RetrievalTransformerVision Language ModelVideoTextMultimodality
🎯 What it does: Proposes the HiTeA framework, utilizing a hierarchical temporal perspective for video-language pre-training.
HiVLP: Hierarchical Interactive Video-Language Pre-Training
Bin Shao (Huawei Noah's Ark Lab), Youliang Yan (Huawei Noah's Ark Lab)
GenerationRetrievalTransformerVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes HiVLP, which utilizes multi-scale visual features and cross-modal interaction with text to perform end-to-end video-text pre-training and unify understanding and generation tasks.
HM-ViT: Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer
Hao Xiang (University of California), Jiaqi Ma (University of California)
Object DetectionAutonomous DrivingTransformerMultimodalityPoint Cloud
🎯 What it does: A heterogeneous multimodal vehicle-to-vehicle collaborative perception framework named HM-ViT is proposed, which can share and fuse information among vehicles with different numbers and types of sensors, enhancing 3D object detection performance.
HMD-NeMo: Online 3D Avatar Motion Generation From Sparse Observations
Sadegh Aliakbarian (Microsoft), Darren Cosker (Microsoft)
GenerationPose EstimationRecurrent Neural NetworkTransformerTime Series
🎯 What it does: This paper presents HMD-NeMo, a unified framework that can generate complete human motion in real-time from 6-DoF signals of HMD head and hand, compatible with both motion controllers and hand tracking scenarios.
Holistic Geometric Feature Learning for Structured Reconstruction
Ziqiong Lu (Wuhan University), Xianwei Zheng (Wuhan University)
Convolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: In the structural reconstruction task, a frequency domain feature learning strategy called F-Learn is introduced, which enhances topological reasoning accuracy by performing frequency domain convolution to fuse geometric fragments on low-level features.
Holistic Label Correction for Noisy Multi-Label Classification
Xiaobo Xia (University of Sydney), Tongliang Liu (University of Sydney)
ClassificationGraph Neural NetworkImage
🎯 What it does: To address the issue of label noise in multi-label classification, a holistic correction (HLC) method utilizing instance-label and label-label dependencies is proposed to achieve automatic correction of noisy labels.
HollowNeRF: Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation
Xiufeng Xie (Oppo Mobile Telecommunications Corporation), Stephen Huang (Oppo Mobile Telecommunications Corporation)
CompressionNeural Radiance FieldPoint Cloud
🎯 What it does: A hashgrid-based NeRF compression method called HollowNeRF has been designed and implemented. During the training phase, a trainable 3D saliency grid and an ADMM trimmer are used to automatically sparsify the feature grid, thereby reducing model parameters.
HoloAssist: an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World
Xin Wang (Microsoft), Marc Pollefeys (ETH Zurich)
Pose EstimationAnomaly DetectionTransformerVideoMultimodalityBenchmark
🎯 What it does: This paper presents the HoloAssist dataset, which records 166 hours of audiovisual and multimodal data from two-person collaborative physical task operations, and establishes three benchmark tasks based on this: error detection, intervention type prediction, and 3D hand pose prediction.
HoloFusion: Towards Photo-realistic 3D Generative Modeling
Animesh Karnewar, David Novotny
GenerationData SynthesisSuper ResolutionDiffusion modelNeural Radiance FieldPoint CloudMesh
🎯 What it does: The HOLOFUSION method is proposed, which achieves high-quality, view-consistent 3D rendering by jointly training a 3D diffusion model and a 2D super-resolution network.
Homeomorphism Alignment for Unsupervised Domain Adaptation
Lihua Zhou (University of Electronic Science and Technology of China), Ferrante Neri (University of Surrey)
Domain AdaptationFlow-based ModelImage
🎯 What it does: A homomorphic mapping (HMA) based on Invertible Neural Networks (INN) is proposed, establishing feature spaces on both the source and target domains, maintaining topological structure through homomorphic mapping and achieving distribution alignment, ultimately training the model simultaneously in both spaces.
Homography Guided Temporal Fusion for Road Line and Marking Segmentation
Shan Wang (Data61 CSIRO), Hongdong Li (Australian National University)
SegmentationAutonomous DrivingImageVideo
🎯 What it does: This paper proposes a homotopy transformation-based temporal fusion module called HomoFusion, which utilizes information from adjacent frames to recover occluded lane lines and markings, achieving lightweight lane marking segmentation.
HopFIR: Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation
Kai Zhai (Hefei University of Technology), Shanlin Yang (Hefei University of Technology)
Pose EstimationGraph Neural NetworkTransformerImage
🎯 What it does: This paper proposes a 3D human pose estimation framework named HopFIR, which mainly consists of two modules: Hop-wise GraphFormer (HGF) and Intragroup Joint Refinement (IJR);
HOSNeRF: Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video
Jia-Wei Liu (National University of Singapore), Mike Zheng Shou (National University of Singapore)
GenerationData SynthesisNeural Radiance FieldOptical FlowVideo
🎯 What it does: This paper presents HOSNeRF, a 360° free viewpoint rendering method based on single-shot video, capable of pausing and rendering dynamic human-object-scene interactions at any moment.
Householder Projector for Unsupervised Latent Semantics Discovery
Yue Song (University of Trento), Wei Wang (Beijing Jiaotong University)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a low-rank orthogonal projector based on Householder transformations (Householder Projector) for reparameterizing the projection matrix of StyleGAN, thereby discovering clearer and more controllable semantic directions in the latent space.
How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability
Zijian Wang (University of Queensland), Mahsa Baktashmotlagh (University of Queensland)
ClassificationImageBenchmark
🎯 What it does: This paper proposes a transferability assessment metric for pre-trained models based on the theory of neural collapse (NCTI) and validates its effectiveness on 11 downstream visual classification tasks.
How Much Temporal Long-Term Context is Needed for Action Segmentation?
Emad Bahrami (University of Bonn), Juergen Gall (University of Bonn)
SegmentationTransformerVideo
🎯 What it does: A temporal action segmentation model named LTContext is proposed, which captures both global and local temporal contexts of videos using sparse attention and local window attention simultaneously.
How to Boost Face Recognition with StyleGAN?
Artem Sevastopolskiy, Matthias Nießner (Technical University of Munich)
RecognitionGenerationConvolutional Neural NetworkSupervised Fine-TuningGenerative Adversarial NetworkImageBenchmark
🎯 What it does: Utilizing a large-scale unlabeled facial image dataset to train a generator using StyleGAN2-ADA, then mapping the images to latent space with a pSp encoder, and finally transferring the encoder's convolutional layers to a facial recognition network for fine-tuning on labeled data, thereby enhancing facial recognition performance.
How to Choose your Best Allies for a Transferable Attack?
Thibault Maho (University of Rennes), Teddy Furon (University of Rennes)
Adversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: A new metric for evaluating the transferability of adversarial attacks is proposed, emphasizing the relationship between distortion and attack success rate (ASR), and based on this, a source model selection method called FiT is introduced.
HRS-Bench: Holistic, Reliable and Scalable Benchmark for Text-to-Image Models
Eslam Mohamed Bakr (King Abdullah University of Science and Technology), Mohamed Elhoseiny (AWS AI Amazon)
Object DetectionGenerationTransformerLarge Language ModelVision Language ModelDiffusion modelImageTextBenchmark
🎯 What it does: A comprehensive evaluation framework named HRS-Bench is proposed to measure 13 capabilities of text-to-image (T2I) models, covering 50 application scenarios, and to conduct automated evaluation and manual verification of 9 existing large T2I models.
HSE: Hybrid Species Embedding for Deep Metric Learning
Bailin Yang (Zhejiang Gongshang University), Chao Song (Zhejiang Gongshang University)
RetrievalContrastive LearningImage
🎯 What it does: In deep metric learning, this paper proposes the Hybrid Species Embedding (HSE) method, which generates unlabeled mixed samples (Hybrid species) through mixed sample data augmentation, serving as additional training signals to enhance the generalization ability of the embedding space.
HSR-Diff: Hyperspectral Image Super-Resolution via Conditional Diffusion Models
Chanyue Wu (Northwestern Polytechnical University), Qiang Shen (Aberystwyth University)
RestorationSuper ResolutionTransformerDiffusion modelImage
🎯 What it does: A super-resolution method HSR-Diff for fusing high-resolution multispectral images with low-resolution hyperspectral images based on a conditional diffusion model is proposed.
HTML: Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation
Mingfei Han (University of Technology Sydney), Yu Qiao (Shanghai AI Laboratory)
Object DetectionSegmentationTransformerVideoTextMultimodality
🎯 What it does: A hybrid temporal scale multimodal learning framework (HTML) is proposed to address the issue of descriptive diversity in open video object segmentation by constructing hierarchical visual-language interactions at different temporal scales.
Human from Blur: Human Pose Tracking from Blurry Images
Yiming Zhao (ETH Zurich), Martin R. Oswald (Microsoft)
Object TrackingPose EstimationOptimizationImage
🎯 What it does: Through differentiable rendering and inverse optimization, recover the 3D human shape, texture, and sub-frame level motion trajectories in blurred images.
Human Part-wise 3D Motion Context Learning for Sign Language Recognition
Taeryung Lee (Seoul National University), Kyoung Mu Lee (Seoul National University)
RecognitionPose EstimationTransformerVideoMultimodality
🎯 What it does: The P3D framework is proposed, utilizing human part-level motion context learning to achieve high-accuracy word-level sign language recognition.
Human Preference Score: Better Aligning Text-to-Image Models with Human Preference
Xiaoshi Wu (Multimedia Laboratory, Chinese University of Hong Kong), Hongsheng Li (SenseTime Research)
GenerationRecommendation SystemTransformerSupervised Fine-TuningDiffusion modelImageText
🎯 What it does: Collected 98,807 Stable Diffusion generated images based on Discord and their human selections, trained a human preference classifier to obtain the Human Preference Score (HPS), and fine-tuned Stable Diffusion using HPS through LoRA to make the generated images more aligned with human aesthetics and intentions.
Human-centric Scene Understanding for 3D Large-scale Scenarios
Yiteng Xu (ShanghaiTech University), Yuexin Ma (ShanghaiTech University)
RecognitionSegmentationTransformerMultimodalityPoint Cloud
🎯 What it does: A large-scale multimodal human-centered scene understanding dataset HuCenLife is proposed, and based on this, a point cloud instance segmentation and action recognition method for human interaction is designed.
Human-Inspired Facial Sketch Synthesis with Dynamic Adaptation
Fei Gao (Xidian University), Nannan Wang (Xidian University)
Image TranslationGenerationData SynthesisConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A facial sketch synthesis method (HIDA) based on 3D geometry, 2D appearance, and global style dynamic adjustment is proposed, capable of generating high-quality sketches with multiple styles that are consistent in style.
HumanMAC: Masked Motion Completion for Human Motion Prediction
Ling-Hao Chen (Tsinghua University), Tongliang Liu (University of Sydney)
GenerationPose EstimationTransformerDiffusion modelVideo
🎯 What it does: A diffusion model called HumanMAC based on masked completion is proposed for human motion prediction.
Humans in 4D: Reconstructing and Tracking Humans with Transformers
Shubham Goel (University of California), Jitendra Malik (University of California)
Object TrackingPose EstimationTransformerVision Language ModelImageVideo
🎯 What it does: A fully transformer architecture HMR 2.0 is proposed for single image recovery of 3D human meshes, and a 4DHumans system is constructed to achieve 3D recovery and tracking of multiple people in videos;
HumanSD: A Native Skeleton-Guided Diffusion Model for Human Image Generation
Xuan Ju (Chinese University of Hong Kong), Qiang Xu (International Digital Economy Academy)
GenerationData SynthesisPose EstimationDiffusion modelImageText
🎯 What it does: We propose HumanSD, a native skeleton-guided diffusion model based on Stable Diffusion, for precise control of human poses and generating multi-scene portraits.
Hybrid Spectral Denoising Transformer with Guided Attention
Zeqiang Lai (Beijing Institute of Technology), Ying Fu (Beijing Institute of Technology)
RestorationTransformerImage
🎯 What it does: A hybrid spectral denoising transformer (HSDT) is proposed for hyperspectral image denoising.
HybridAugment++: Unified Frequency Spectra Perturbations for Model Robustness
Mehmet Kerim Yucel (Hacettepe University), Pinar Duygulu (Hacettepe University)
ClassificationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: Two unsupervised data augmentation methods based on the spectrum, HybridAugment and HybridAugment++, are proposed. By mixing high and low frequencies and amplitude-phase information of images during training, the CNN is forced to rely more on low-frequency and phase features, thereby enhancing the model's robustness to distribution shifts.
Hyperbolic Audio-visual Zero-shot Learning
Jie Hong (Australian National University), Lars Petersson (Monash University)
ClassificationRecognitionContrastive LearningVideoMultimodalityAudio
🎯 What it does: An alignment loss for audio-visual zero-shot learning in hyperbolic space is proposed, and cross-modal feature alignment and hierarchical mining are achieved through three alignment modules.
Hyperbolic Chamfer Distance for Point Cloud Completion
Fangzhou Lin (Worcester Polytechnic Institute), Ziming Zhang (Worcester Polytechnic Institute)
RestorationGenerationPoint Cloud
🎯 What it does: This paper proposes the HyperCD, a Chamfer distance defined in hyperbolic space, for point cloud completion tasks and validates its effectiveness across various network architectures.
HyperDiffusion: Generating Implicit Neural Fields with Weight-Space Diffusion
Ziya Erkoç, Angela Dai (Technical University of Munich)
GenerationData SynthesisTransformerDiffusion modelNeural Radiance FieldPoint CloudMesh
🎯 What it does: A new method called HyperDiffusion is proposed for directly generating diffusion models on neural implicit fields (MLP weights) to unconditionally generate high-quality 3D shapes and 4D animation sequences.
HyperReenact: One-Shot Reenactment via Jointly Learning to Refine and Retarget Faces
Stella Bounareli (Kingston University), Georgios Tzimiropoulos (Queen Mary University)
Image TranslationGenerationGenerative Adversarial NetworkImageVideo
🎯 What it does: This paper proposes a single-frame speaker reenactment method called HyperReenact, based on StyleGAN2 and super networks, which can generate high-quality speaker avatars without noticeable visual artifacts under extreme head pose variations and cross-person scenarios.
I Can't Believe There's No Images! Learning Visual Tasks Using only Language Supervision
Sophia Gu (Allen Institute for Artificial Intelligence), Aniruddha Kembhavi (Allen Institute for Artificial Intelligence)
Domain AdaptationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a cross-modal zero-shot transfer method (CLOSE), which trains a model using text data in a visual-text joint embedding space obtained through contrastive learning, and then replaces text embeddings with image embeddings to complete visual tasks. It demonstrates that training solely on text can achieve performance close to that of image-trained models in tasks such as image captioning, visual entailment, VQA, and visual news.
I-ViT: Integer-only Quantization for Efficient Vision Transformer Inference
Zhikai Li (Institute of Automation Chinese Academy of Sciences), Qingyi Gu (Institute of Automation Chinese Academy of Sciences)
ClassificationComputational EfficiencyTransformerImage
🎯 What it does: Proposes I-ViT, a full quantization scheme that uses only integer operations during the inference process of Vision Transformer.