CVPR 2023 Papers — Page 2
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2353 papers
Adversarial Normalization: I Can Visualize Everything (ICE)
Hoyoung Choi (Hanyang University), Kyungsik Han (Hanyang University)
SegmentationExplainability and InterpretabilityAdversarial AttackTransformerImage
🎯 What it does: Proposes a method to visualize the interpretability of visual Transformers (ICE) by classifying each image patch and using adversarial normalization.
Adversarial Robustness via Random Projection Filters
Minjing Dong (University of Sydney), Chang Xu (University of Sydney)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A Random Projection Filters (RPF) scheme is proposed, where part of the convolutional kernels is replaced with random Gaussian weights, and adversarial training is conducted based on this to enhance the network's robustness against both white-box and black-box attacks.
Adversarially Masking Synthetic To Mimic Real: Adaptive Noise Injection for Point Cloud Segmentation Adaptation
Guangrui Li (University of Technology Sydney), Yi Yang (Zhejiang University)
SegmentationDomain AdaptationAutonomous DrivingGenerative Adversarial NetworkPoint Cloud
🎯 What it does: To address the semantic segmentation transfer from synthetic point clouds to real point clouds, adaptation is achieved through dynamic masking of the source point clouds.
Adversarially Robust Neural Architecture Search for Graph Neural Networks
Beini Xie (Tsinghua University), Wenwu Zhu (Tsinghua University)
Adversarial AttackNeural Architecture SearchGraph Neural NetworkGraph
🎯 What it does: A robust neural architecture search framework for graph neural networks, G-RNA, is proposed. By incorporating defense operations such as graph structure masks into the search space and defining robustness metrics, it automatically finds attack-resistant GNN architectures using single-path one-shot NAS and evolutionary search.
AeDet: Azimuth-Invariant Multi-View 3D Object Detection
Chengjian Feng (Meituan Inc), Lin Ma (Meituan Inc)
Object DetectionAutonomous DrivingConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a multi-view 3D object detection framework called AeDet, which achieves azimuth-invariance and specifically models the radial symmetry of multi-camera BEV features, unifying the predicted objects under different azimuth angles.
Affection: Learning Affective Explanations for Real-World Visual Data
Panos Achlioptas (Snap Inc), Sergey Tulyakov (Snap Inc)
GenerationTransformerImageText
🎯 What it does: This paper presents a large-scale 'Affection' dataset, which contains 85,007 real-world images, 526,749 emotional responses, and corresponding free-text explanations, and based on this, trains an emotion-driven and conversational visual-text generation model.
Affordance Diffusion: Synthesizing Hand-Object Interactions
Yufei Ye (Carnegie Mellon University), Sifei Liu (NVIDIA)
GenerationData SynthesisPose EstimationDiffusion modelScore-based ModelImage
🎯 What it does: This paper proposes a two-step generative framework based on diffusion models (LayoutNet + ContentNet) that can synthesize interactive hand poses and hand-object interaction images from a single RGB image of an object, and directly extract 3D hand poses from them.
Affordance Grounding From Demonstration Video To Target Image
Joya Chen (National University of Singapore), Mike Zheng Shou (National University of Singapore)
Object DetectionSegmentationTransformerImageVideo
🎯 What it does: Proposes two methods, Affordance Transformer (Afformer) and Masked Affordance Hand (MaskAHand), to achieve fine-grained interaction (affordance) localization from video to image.
Affordances From Human Videos as a Versatile Representation for Robotics
Shikhar Bahl (Carnegie Mellon University), Deepak Pathak (Carnegie Mellon University)
Robotic IntelligenceConvolutional Neural NetworkTransformerReinforcement LearningVideoMultimodality
🎯 What it does: This paper proposes the Vision-Robotics Bridge (VRB), which constructs a robot-friendly operable visual empowerment model by learning contact points and subsequent trajectories from human interaction videos, and seamlessly integrates it into four robotic learning paradigms (offline imitation learning, reward-free exploration, goal-driven learning, and action space parameterization), significantly improving learning efficiency in real environments.
AGAIN: Adversarial Training With Attribution Span Enlargement and Hybrid Feature Fusion
Shenglin Yin (Peking University), Zhen Xiao (Peking University)
OptimizationAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes an adversarial training method based on extended attribution span and mixed feature fusion (AGAIN) to enhance the generalization performance of robust models.
Alias-Free Convnets: Fractional Shift Invariance via Polynomial Activations
Hagay Michaeli (Technion), Daniel Soudry (Technion)
ClassificationRecognitionConvolutional Neural NetworkImage
🎯 What it does: A fully aliasing-free convolutional network (AFC) is proposed, achieving invariance to integer and sub-pixel translations by rewriting modules such as downsampling, activation functions, and normalization within the convolutional network.
Align and Attend: Multimodal Summarization With Dual Contrastive Losses
Bo He (University of Maryland), Zhaowen Wang (Adobe Research)
TransformerMixture of ExpertsContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes a unified Transformer framework A2Summ for multimodal summarization, combining temporal alignment and dual contrastive loss;
Align Your Latents: High-Resolution Video Synthesis With Latent Diffusion Models
Andreas Blattmann (LMU Munich), Karsten Kreis (NVIDIA)
GenerationData SynthesisAutonomous DrivingDiffusion modelVideoText
🎯 What it does: This paper proposes the Video Latent Diffusion Model (Video LDM), which achieves high-resolution video generation by inserting a temporal alignment layer into a pre-trained image LDM, and further fine-tunes the upsampler for temporal adjustments, supporting long videos, text-to-video, and personalized video generation.
AligNeRF: High-Fidelity Neural Radiance Fields via Alignment-Aware Training
Yifan Jiang (University of Texas at Austin), Tianfan Xue (Google Research)
GenerationData SynthesisConvolutional Neural NetworkNeural Radiance FieldOptical FlowImage
🎯 What it does: Proposes AligNeRF, which enhances rendering quality in high-resolution NeRF training through alignment-aware training, convolutional auxiliary networks, and high-frequency loss.
Aligning Bag of Regions for Open-Vocabulary Object Detection
Size Wu (Nanyang Technological University), Chen Change Loy (The University of Hong Kong)
Object DetectionConvolutional Neural NetworkVision Language ModelContrastive LearningImageText
🎯 What it does: The BARON method is proposed, which extends open vocabulary object detection from single region alignment to 'region pack' alignment, utilizing a combined structure of pre-trained visual language models to enhance the performance of new category detection.
Aligning Step-by-Step Instructional Diagrams to Video Demonstrations
Jiahao Zhang (Australian National University), Stephen Gould (Australian National University)
RetrievalRepresentation LearningConvolutional Neural NetworkContrastive LearningVideoMultimodality
🎯 What it does: A multimodal model based on contrastive learning is proposed to align furniture assembly video segments with the illustrated steps in their instruction manuals.
All Are Worth Words: A ViT Backbone for Diffusion Models
Fan Bao (Tsinghua University), Jun Zhu (Tsinghua University)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: This paper proposes a U-ViT architecture based on Vision Transformer (ViT) for noise prediction in diffusion models.
All in One: Exploring Unified Video-Language Pre-Training
Jinpeng Wang (National University of Singapore), Mike Zheng Shou (National University of Singapore)
RecognitionRetrievalTransformerVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: A unified end-to-end video language pre-training model called All-in-one Transformer is proposed, which learns multimodal representations directly from raw video pixels and a text encoder without the need for separate visual/language encoders.
All-in-Focus Imaging From Event Focal Stack
Hanyue Lou (Peking University), Boxin Shi (Peking University)
RestorationDepth EstimationConvolutional Neural NetworkImage
🎯 What it does: An algorithm is proposed that combines the event stack (EFS) recorded during continuous focal length scanning with a single defocused RGB image to generate all-in-focus images.
All-in-One Image Restoration for Unknown Degradations Using Adaptive Discriminative Filters for Specific Degradations
Dongwon Park (Seoul National University), Se Young Chun (Seoul National University)
RestorationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A global image restoration model ADMS is proposed for unknown multiple degradations, which adds about 3% of dedicated filters sparsely for each degradation in a unified network and dynamically activates them through a degradation classifier.
ALOFT: A Lightweight MLP-Like Architecture With Dynamic Low-Frequency Transform for Domain Generalization
Jintao Guo (Nanjing University), Yinghuan Shi (Nanjing University)
ClassificationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: A lightweight MLP-like architecture is proposed, introducing Dynamic Low-Frequency Spectrum Transformation (ALOFT) to enhance domain generalization performance.
ALSO: Automotive Lidar Self-Supervision by Occupancy Estimation
Alexandre Boulch (Valeo), Renaud Marlet (Valeo)
Object DetectionSegmentationAutonomous DrivingContrastive LearningPoint Cloud
🎯 What it does: A self-supervised pre-training method based on the visibility information of LiDAR point clouds is proposed: by constructing query points on the front, back, and sight segments, the occupancy reconstruction task is used to train the point cloud backbone network, and the resulting latent vectors are then used for downstream tasks such as semantic segmentation and object detection.
AltFreezing for More General Video Face Forgery Detection
Zhendong Wang (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
ClassificationRecognitionConvolutional Neural NetworkVideo
🎯 What it does: This paper proposes a method to simultaneously capture spatial and temporal forgery traces in video facial forgery within a single model. The core of the method is the AltFreezing training strategy, which alternately freezes spatial and temporal convolution weights, combined with video-level data augmentation to enhance the generalization ability of facial forgery detection.
ALTO: Alternating Latent Topologies for Implicit 3D Reconstruction
Zhen Wang (University of California), Achuta Kadambi (University of California)
GenerationData SynthesisComputational EfficiencyConvolutional Neural NetworkNeural Radiance FieldPoint CloudMesh
🎯 What it does: An ALTO framework that alternates between point and grid latent spaces is proposed for high-fidelity implicit 3D reconstruction.
Ambiguity-Resistant Semi-Supervised Learning for Dense Object Detection
Chang Liu (Shanghai University), Jingdong Wang (Baidu Inc)
Object DetectionSupervised Fine-TuningImage
🎯 What it does: This paper addresses the issues of selection ambiguity and assignment ambiguity in single-stage object detectors within semi-supervised learning, proposing a framework named ARSL to alleviate these two types of problems, thereby significantly improving detection performance.
Ambiguous Medical Image Segmentation Using Diffusion Models
Aimon Rahman (Johns Hopkins University), Vishal M. Patel (Johns Hopkins University)
SegmentationDiffusion modelImageBiomedical DataMagnetic Resonance ImagingComputed TomographyUltrasound
🎯 What it does: A medical image fuzzy segmentation method based on diffusion models (CIMD) is proposed, which can generate multiple feasible segmentation results from a single image;
AMT: All-Pairs Multi-Field Transforms for Efficient Frame Interpolation
Zhen Li (Nankai University), Ming-Ming Cheng (Nankai University)
RestorationComputational EfficiencyConvolutional Neural NetworkOptical FlowVideo
🎯 What it does: This paper proposes the AMT (All-Pairs Multi-Field Transforms) network for efficient video frame interpolation, utilizing bidirectional all-pairs correlation fields and multi-field refinement to achieve more accurate task-oriented optical flow.
An Actor-Centric Causality Graph for Asynchronous Temporal Inference in Group Activity
Zhao Xie (Hefei University of Technology), Jiao Chang (Hefei University of Technology)
RecognitionPose EstimationGraph Neural NetworkTransformerVideo
🎯 What it does: This study investigates asynchronous causal relationships in group activity recognition and proposes the Actor-Centric Causal Graph (ACCG) model to detect and integrate asynchronous causal relationships with synchronous spatiotemporal relationships.
An Empirical Study of End-to-End Video-Language Transformers With Masked Visual Modeling
Tsu-Jui Fu (University of California Santa Barbara), Zicheng Liu (Microsoft)
RecognitionRetrievalTransformerVision Language ModelVideoText
🎯 What it does: This paper proposes an improved version of VIOLETv2 by systematically evaluating eight types of Masked Visual Modeling (MVM) objectives in conjunction with the video-language Transformer VIOLET, achieving significant improvements on 13 VidL benchmarks.
An Erudite Fine-Grained Visual Classification Model
Dongliang Chang (Beijing University of Posts and Telecommunications), Zhanyu Ma (Beijing University of Posts and Telecommunications)
ClassificationMeta LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: Build a single model to jointly train and identify fine-grained labels from multiple fine-grained visual classification datasets without relying on coarse-grained labels.
An Image Quality Assessment Dataset for Portraits
Nicolas Chahine (DXOMARK), Jean Ponce (Institute of Mathematical Sciences and Center for Data Science New York University)
ClassificationData-Centric LearningConvolutional Neural NetworkImageBenchmark
🎯 What it does: A quality assessment dataset specifically for portrait photos, PIQ23, has been constructed and made public, and statistical analysis has been conducted based on paired scores obtained from expert comparative experiments; on this basis, a semantic-aware deep BIQA model, SEM-HyperIQA, has been proposed that can adapt to quality scales in different scenarios.
An In-Depth Exploration of Person Re-Identification and Gait Recognition in Cloth-Changing Conditions
Weijia Li (Beijing Jiaotong University), Yao Zhao (Beijing Jiaotong University)
RecognitionRetrievalVideoBenchmark
🎯 What it does: A new video ReID and gait recognition benchmark, CCPG, with seven types of clothing variations has been constructed, and methods for video ReID and gait recognition have been compared on it.
Analyzing and Diagnosing Pose Estimation With Attributions
Qiyuan He (National University of Singapore), Angela Yao (National University of Singapore)
Pose EstimationExplainability and InterpretabilityGraph Neural NetworkImage
🎯 What it does: This paper proposes Pose Integrated Gradient (PoseIG), a pixel-level attribution method for pose estimation.
Analyzing Physical Impacts Using Transient Surface Wave Imaging
Tianyuan Zhang (Carnegie Mellon University), Srinivasa G. Narasimhan (Carnegie Mellon University)
Time SeriesPhysics Related
🎯 What it does: Capture transient surface waves using a dual-shutter laser speckle camera with sparse measurement points, locate the impact source position using an elastic wave propagation model, and explore the inference of impact intensity and object properties.
Anchor3DLane: Learning To Regress 3D Anchors for Monocular 3D Lane Detection
Shaofei Huang (Institute of Information Engineering, Chinese Academy of Sciences), Si Liu (Institute of Artificial Intelligence, Beihang University)
Object DetectionAutonomous DrivingConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes Anchor3DLane, a complete end-to-end detection framework that directly regresses 3D lanes through 3D line anchors in front view (FV) features without the need for BEV transformation.
AnchorFormer: Point Cloud Completion From Discriminative Nodes
Zhikai Chen (University of Science and Technology of China), Tao Mei (HiDream.ai Inc.)
RestorationGenerationAutonomous DrivingTransformerPoint Cloud
🎯 What it does: Proposes AnchorFormer, which utilizes learned anchor points and point shape transformations to achieve point cloud completion.
ANetQA: A Large-Scale Benchmark for Fine-Grained Compositional Reasoning Over Untrimmed Videos
Zhou Yu (Hangzhou Dianzi University), Jun Yu (Lenovo)
RetrievalTransformerVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: The ANetQA video question answering benchmark is proposed, which constructs fine-grained scene graphs based on ActivityNet long videos and automatically generates 1.4 billion question-answer pairs.
Angelic Patches for Improving Third-Party Object Detector Performance
Wenwen Si (University of Pennsylvania), Osbert Bastani (University of Pennsylvania)
Object DetectionAdversarial AttackImageVideo
🎯 What it does: For third-party object detection models, an 'angelic patch' is proposed, which enhances detection robustness and accuracy by attaching specific textures to the target object.
Annealing-Based Label-Transfer Learning for Open World Object Detection
Yuqing Ma (Beihang University), Xianglong Liu (Beihang University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: A label transfer framework based on sawtooth annealing scheduling (Annealing-based Label-Transfer) is proposed, enabling the simultaneous learning of known and unknown categories in open-world object detection (OWOD) tasks without manually selecting unknown boxes.
AnyFlow: Arbitrary Scale Optical Flow With Implicit Neural Representation
Hyunyoung Jung (Seoul National University), Denis Demandolx (Meta Reality Labs)
GenerationOptical FlowImageVideo
🎯 What it does: This paper presents AnyFlow, a network that models optical flow as a continuous coordinate function, capable of generating high-quality optical flow at arbitrary scales from low-resolution inputs.
Architectural Backdoors in Neural Networks
Mikel Bober-Irizar (University of Cambridge), Nicolas Papernot (University of Toronto)
Adversarial AttackNeural Architecture SearchConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a new type of backdoor attack, namely Model Architecture Backdoor (MAB), which embeds trigger detection logic within the network structure itself, making the backdoor independent of weights and capable of retaining functionality even after the model is retrained from scratch.
Architecture, Dataset and Model-Scale Agnostic Data-Free Meta-Learning
Zixuan Hu (Tsinghua University), Dacheng Tao (JD Explore Academy)
ClassificationMeta LearningContrastive LearningImage
🎯 What it does: A data-free meta-learning framework called PURER is proposed, which utilizes the knowledge from pre-trained models for task learning and is applicable to models of different architectures, scales, and datasets.
ARCTIC: A Dataset for Dexterous Bimanual Hand-Object Manipulation
Zicong Fan (ETH Zurich), Otmar Hilliges (ETH Zurich)
Pose EstimationRobotic IntelligenceRecurrent Neural NetworkVideoMeshBenchmark
🎯 What it does: The ARCTIC dataset is proposed, which includes videos of two-handed manipulation of articulated objects and high-quality 3D hand and object meshes, along with the design of two benchmark tasks: consistent motion reconstruction and interaction field estimation.
Are Binary Annotations Sufficient? Video Moment Retrieval via Hierarchical Uncertainty-Based Active Learning
Wei Ji (National University of Singapore), Tat-seng Chua (National University of Singapore)
RetrievalConvolutional Neural NetworkSupervised Fine-TuningVideo
🎯 What it does: This paper proposes an interactive hierarchical uncertainty active learning framework (HUAL) that uses binary labels instead of complete timestamps for video moment retrieval training, significantly reducing annotation costs.
Are Data-Driven Explanations Robust Against Out-of-Distribution Data?
Tang Li (University of Delaware), Xi Peng (University of Delaware)
Domain AdaptationExplainability and InterpretabilityImage
🎯 What it does: This paper studies the robustness of data-driven explanations on out-of-distribution (OOD) data and proposes an unsupervised distributed robust explanation framework (DRE) based on distribution consistency and sparsity constraints.
Are Deep Neural Networks SMARTer Than Second Graders?
Anoop Cherian (Mitsubishi Electric Research Labs), Joshua B. Tenenbaum (Massachusetts Institute of Technology)
Meta LearningTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: Proposed the SMART task and the SMART-101 dataset to evaluate the performance of deep neural networks in children's algorithmic reasoning.
Are We Ready for Vision-Centric Driving Streaming Perception? The ASAP Benchmark
Xiaofeng Wang (Institute of Automation, Chinese Academy of Sciences), Xingang Wang (Institute of Automation, Chinese Academy of Sciences)
Object DetectionAutonomous DrivingMultimodalityBenchmark
🎯 What it does: The ASAP benchmark is proposed to evaluate the performance of camera-based 3D detection in self-driving scenarios under real-time inference (streaming) conditions, and it is extended to 12Hz high frame rate annotations on the nuScenes dataset. Additionally, the SPUR evaluation protocol, speed prediction, and learning-based prediction baseline methods are designed to compensate for inference delays. Comprehensive comparative experiments are conducted on various existing 3D detection models under different GPU and shared resource environments.
ARKitTrack: A New Diverse Dataset for Tracking Using Mobile RGB-D Data
Haojie Zhao (Dalian University of Technology), Huchuan Lu (Dalian University of Technology)
Object TrackingSegmentationTransformerSimultaneous Localization and MappingVideoBenchmark
🎯 What it does: This paper presents a dataset called ARKitTrack, which consists of 300 RGB-D video sequences (a total of 455 targets and 229.7K frames) collected using the iPhone LiDAR camera. It provides axis-aligned bounding boxes, pixel-level masks, frame-level attributes, and camera intrinsic parameters and poses. Additionally, a unified RGB-D tracking baseline is provided, achieving a closed loop between VOT and VOS.
ARO-Net: Learning Implicit Fields From Anchored Radial Observations
Yizhi Wang (Shenzhen University), Ruizhen Hu (Reichman University)
GenerationRepresentation LearningTransformerPoint Cloud
🎯 What it does: Proposes a query-specific shape encoding based on Anchored Radial Observations (ARO) for learning implicit field representations of 3D shapes;
AShapeFormer: Semantics-Guided Object-Level Active Shape Encoding for 3D Object Detection via Transformers
Zechuan Li (Hunan University), Naveed Akhtar (The University of Western Australia)
Object DetectionTransformerPoint Cloud
🎯 What it does: A pluggable AShapeFormer module is proposed to actively encode object shape information in 3D object detection, thereby improving the quality of candidate points and detection accuracy.
ASPnet: Action Segmentation With Shared-Private Representation of Multiple Data Sources
Beatrice van Amsterdam (Wellcome EPSRC Centre for Interventional and Surgical Sciences), Danail Stoyanov (Medtronic)
RecognitionSegmentationTransformerOptical FlowVideoMultimodality
🎯 What it does: This paper proposes ASPnet, a multi-source data fusion action segmentation network that can achieve precise temporal segmentation on multimodal data such as video, accelerometer, and optical flow.
AssemblyHands: Towards Egocentric Activity Understanding via 3D Hand Pose Estimation
Takehiko Ohkawa (Meta Reality Labs), Cem Keskin (Meta Reality Labs)
RecognitionPose EstimationConvolutional Neural NetworkGraph Neural NetworkImage
🎯 What it does: Constructed the AssemblyHands dataset, providing high-quality 3D hand pose annotations for 3.0M images, particularly focusing on hand-object interactions from an egocentric perspective.
AstroNet: When Astrocyte Meets Artificial Neural Network
Mengqiao Han (Beijing Institute of Technology), Xiabi Liu (Beijing Institute of Technology)
ClassificationOptimizationNeural Architecture SearchImage
🎯 What it does: This paper studies the impact of astrocytes on the regulation of neural network connections, proposing AstroNet, which achieves adaptive connection regulation by introducing the Astrocyte-Neuron model, thereby efficiently optimizing the network structure without compromising performance.
AsyFOD: An Asymmetric Adaptation Paradigm for Few-Shot Domain Adaptive Object Detection
Yipeng Gao (Sun Yat-sen University), Wei-Shi Zheng (Pengcheng Lab)
Object DetectionDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: An asymmetric adaptation framework AsyFOD is proposed for few-shot domain adaptive object detection.
Asymmetric Feature Fusion for Image Retrieval
Hui Wu (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
RetrievalTransformerImage
🎯 What it does: Proposes a method in heterogeneous retrieval systems that fuses various global and local features only on the gallery side, using a dynamic mixer to compress them into a single vector, while the query side only employs a lightweight model;
Attention-Based Point Cloud Edge Sampling
Chengzhi Wu (Karlsruhe Institute of Technology), Jürgen Beyerer (Fraunhofer IOSB)
ClassificationSegmentationPoint Cloud
🎯 What it does: A point cloud edge sampling method based on attention (APES) is proposed, which directly selects edge points by calculating the standard deviation/column sum of local or global attention relevance, completing the downsampling of point clouds;
AttentionShift: Iteratively Estimated Part-Based Attention Map for Pointly Supervised Instance Segmentation
Mingxiang Liao (University of Chinese Academy of Sciences), Fang Wan (University of Chinese Academy of Sciences)
Object DetectionSegmentationTransformerImage
🎯 What it does: Proposes the AttentionShift method, which utilizes the translation of token queries and key points in the feature space of ViT to iteratively estimate and optimize part-based attention maps, thereby achieving instance segmentation under single-point supervision.
Attribute-Preserving Face Dataset Anonymization via Latent Code Optimization
Simone Barattin (University of Trento), Nicu Sebe (University of Trento)
RecognitionGenerationSafty and PrivacyGenerative Adversarial NetworkImage
🎯 What it does: By directly optimizing the latent codes in the pre-trained StyleGAN2 latent space, identity de-identification of facial images is achieved while maintaining the facial attributes of the original images.
AttriCLIP: A Non-Incremental Learner for Incremental Knowledge Learning
Runqi Wang (Beihang University), Baochang Zhang (Beihang University)
ClassificationRecognitionTransformerVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes a continual learning framework called AttriCLIP based on CLIP, which fine-tunes only the learnable attribute word vectors through fixed visual and text encoders. It utilizes an attribute word list to select and update text prompts, achieving continuous knowledge learning without the need to expand classifiers or store replay data.
Audio-Visual Grouping Network for Sound Localization From Mixtures
Shentong Mo (Carnegie Mellon University), Yapeng Tian (University of Texas at Dallas)
Object DetectionRepresentation LearningTransformerVideoMultimodalityAudio
🎯 What it does: Designed and implemented an Audio-Video Group Network (AVGN) that can learn category-level semantic features from mixed audio and images, and directly locate the positions of multiple sound sources.
Augmentation Matters: A Simple-Yet-Effective Approach to Semi-Supervised Semantic Segmentation
Zhen Zhao (University of Sydney), Jingdong Wang (Baidu Inc.)
SegmentationKnowledge DistillationImage
🎯 What it does: This paper studies a semi-supervised semantic segmentation method called AugSeg based on a teacher-student framework, focusing on improving data augmentation strategies.
AUNet: Learning Relations Between Action Units for Face Forgery Detection
Weiming Bai (Chinese Academy of Sciences), Weiming Hu (Chinese Academy of Sciences)
Anomaly DetectionTransformerImageVideo
🎯 What it does: A deepfake detection framework based on learning the relationships of facial action units (AUs) is proposed, which includes the Action Units Relation Transformer (ART) and Tampered AU Prediction (TAP). It enhances detection performance by focusing on the relationships of AU regions and local tampering.
Auto-CARD: Efficient and Robust Codec Avatar Driving for Real-Time Mobile Telepresence
Yonggan Fu (Georgia Institute of Technology), Yingyan (Celine) Lin (Georgia Institute of Technology)
CompressionComputational EfficiencyNeural Architecture SearchVideo
🎯 What it does: The Auto-CARD framework is proposed, achieving real-time and robust Codec Avatar driving using only local computing resources from AR/VR devices.
AutoAD: Movie Description in Context
Tengda Han (University of Oxford), Andrew Zisserman (University of Oxford)
GenerationTransformerLarge Language ModelPrompt EngineeringVideoTextMultimodalityAudio
🎯 What it does: This study proposes the AutoAD model, which can automatically generate audio descriptions (AD) that meet the needs of visually impaired individuals from movie clips, specifically generating narrative text within a given time interval.
AutoFocusFormer: Image Segmentation off the Grid
Chen Ziwen (Oregon State University), Li Fuxin (Apple Inc)
ClassificationSegmentationTransformerImagePoint Cloud
🎯 What it does: Proposes AutoFocusFormer, an end-to-end segmentation network that utilizes adaptive downsampling and local attention to process non-grid images.
AutoLabel: CLIP-Based Framework for Open-Set Video Domain Adaptation
Giacomo Zara (University of Trento), Elisa Ricci (University of Trento)
ClassificationDomain AdaptationTransformerVision Language ModelContrastive LearningVideo
🎯 What it does: Proposes the AutoLabel framework, which automatically generates candidate unknown category labels for the target domain on the open-source CLIP, thereby achieving open-set domain adaptation for video action classification in the unlabeled target domain.
Automatic High Resolution Wire Segmentation and Removal
Mang Tik Chiu (University of Illinois Urbana-Champaign), Humphrey Shi (Adobe)
RestorationSegmentationConvolutional Neural NetworkTransformerImageBenchmark
🎯 What it does: This paper proposes a fully automated wire segmentation and removal system for high-resolution images, which includes a two-stage coarse-to-fine segmentation network and a LaMa-based tiled image restoration pipeline.
Autonomous Manipulation Learning for Similar Deformable Objects via Only One Demonstration
Yu Ren (Chinese Academy of Sciences), Yang Cong (Chinese Academy of Sciences)
Robotic IntelligencePoint Cloud
🎯 What it does: A framework for grasping and wearing that can learn from a single demonstration and generalize to deformed objects of the same category is proposed.
AutoRecon: Automated 3D Object Discovery and Reconstruction
Yuang Wang (Zhejiang University), Xiaowei Zhou (Zhejiang University)
Object DetectionSegmentationTransformerNeural Radiance FieldPoint Cloud
🎯 What it does: This paper proposes a fully automated 3D object discovery and reconstruction framework called AutoRecon, which can automatically identify foreground objects from multi-view images and generate background-free 3D models and high-quality 2D segmentations without any manual annotations.
Autoregressive Visual Tracking
Xing Wei (Xi'an Jiaotong University), Yihong Gong (Xi'an Jiaotong University)
Object TrackingTransformerVideo
🎯 What it does: ARTrack is proposed, an end-to-end autoregressive visual tracking framework that treats target trajectories as discrete coordinate sequences, directly decoded through a Transformer without the need for traditional localization heads or post-processing.
Avatars Grow Legs: Generating Smooth Human Motion From Sparse Tracking Inputs With Diffusion Model
Yuming Du (Meta AI), Artsiom Sanakoyeu (Meta AI)
GenerationPose EstimationDiffusion modelVideo
🎯 What it does: A full-body motion synthesis method called AGRoL based on conditional diffusion models has been designed and implemented, which can predict complete human motion using only the 3D pose information of the head and hands from a head-mounted display (HMD).
AVFace: Towards Detailed Audio-Visual 4D Face Reconstruction
Aggelina Chatziagapi (Stony Brook University), Dimitris Samaras (Stony Brook University)
RestorationGenerationTransformerVideoMultimodalityAudio
🎯 What it does: An AVFace method is proposed, utilizing monocular video and audio for 4D face reconstruction without 3D annotations, capable of recovering detailed facial geometry and lip movements.
AVFormer: Injecting Vision Into Frozen Speech Models for Zero-Shot AV-ASR
Paul Hongsuck Seo (Google Research), Cordelia Schmid (Google Research)
RecognitionRecurrent Neural NetworkTransformerSupervised Fine-TuningContrastive LearningVideoMultimodalityAudio
🎯 What it does: Under the premise of keeping the audio ASR model frozen, zero-shot multimodal speech recognition is achieved by injecting CLIP visual features using lightweight adapters and visual projection layers.
Azimuth Super-Resolution for FMCW Radar in Autonomous Driving
Yu-Jhe Li (Carnegie Mellon University), Kris Kitani (Carnegie Mellon University)
Object DetectionSuper ResolutionAutonomous DrivingConvolutional Neural NetworkGenerative Adversarial NetworkPoint Cloud
🎯 What it does: A deep learning-based ADC-SR and Hybrid-SR model has been designed and implemented to achieve azimuth super-resolution in MIMO FMCW radar by predicting missing receiver ADC signals, which can then directly generate high-resolution RAD images and enhance target detection performance.
B-Spline Texture Coefficients Estimator for Screen Content Image Super-Resolution
Byeonghyun Pak (Daegu Gyeongbuk Institute of Science and Technology), Kyong Hwan Jin (Daegu Gyeongbuk Institute of Science and Technology)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an arbitrary scale screen content image super-resolution method based on a B-spline texture coefficient estimator (BTC).
BAAM: Monocular 3D Pose and Shape Reconstruction With Bi-Contextual Attention Module and Attention-Guided Modeling
Hyo-Jun Lee (Chungnam National University), Yeong Jun Koh (Chungnam National University)
Object DetectionPose EstimationAutonomous DrivingConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a monocular 3D vehicle pose and shape reconstruction framework based on Bi-Contextual Attention and Attention-Guided Modeling (BAAM), and introduces 3D Non-Maximum Suppression (3D NMS) to eliminate false detections;
Back to the Source: Diffusion-Driven Adaptation To Test-Time Corruption
Jin Gao (Shanghai Jiao Tong University), Dequan Wang (Shanghai Artificial Intelligence Laboratory)
ClassificationDomain AdaptationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: A method is proposed to project damaged target images back to the source domain during inference using a diffusion model, and then directly use the source domain classifier for prediction.
Backdoor Attacks Against Deep Image Compression via Adaptive Frequency Trigger
Yi Yu (Nanyang Technological University), Alex C. Kot (Nanyang Technological University)
RecognitionSegmentationCompressionAdversarial AttackImageVideo
🎯 What it does: This paper proposes a method to implement multi-target backdoor attacks on learning-based image compression models, utilizing adaptive trigger injection based on the DCT frequency domain, and modifying only the encoder parameters to activate hidden malicious behaviors.
Backdoor Cleansing With Unlabeled Data
Lu Pang (Stony Brook University), Chao Chen (Stony Brook University)
Knowledge DistillationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A backdoor cleaning framework utilizing unlabeled data (which can be in-distribution or out-of-distribution) is proposed: first, the student network is initialized with adaptive hierarchical random resets, and then trained on unlabeled data through knowledge distillation (targeting the soft logits of the teacher network) to remove the backdoor while maintaining normal prediction performance.
Backdoor Defense via Adaptively Splitting Poisoned Dataset
Kuofeng Gao (Tsinghua University), Shu-Tao Xia (Peng Cheng Laboratory)
OptimizationAdversarial AttackData-Centric LearningMeta LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a method to defend against backdoor attacks during the training phase by adaptively partitioning the dataset (ASD), dividing the training set into a clean pool and a contaminated pool, and jointly training the model on both using semi-supervised learning.
Backdoor Defense via Deconfounded Representation Learning
Zaixi Zhang (University of Science and Technology of China), Qingyong Hu (Hong Kong University of Science and Technology)
Anomaly DetectionRepresentation LearningAdversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a causal inference-based backdoor defense framework (CBD) that directly trains a clean model without backdoors from contaminated datasets by learning deconfounded representations.
BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields
Peng Wang (Zhejiang University), Peidong Liu (Westlake University)
RestorationData SynthesisOptimizationNeural Radiance FieldImage
🎯 What it does: This paper proposes a motion-blurred image combined bundle-adjustment neural radiance field (BAD-NeRF), which can simultaneously learn 3D scene representation and the motion trajectory during camera exposure, thus achieving deblurring and high-quality view synthesis under severely blurred images.
BAEFormer: Bi-Directional and Early Interaction Transformers for Bird's Eye View Semantic Segmentation
Cong Pan (National Laboratory of Pattern Recognition), Zhaoxiang Zhang (National Laboratory of Pattern Recognition)
SegmentationAutonomous DrivingTransformerImage
🎯 What it does: This paper proposes a Transformer-based bird's-eye view semantic segmentation framework called BAEFormer, which can directly convert perspective images from multi-view cameras into BEV space for semantic segmentation.
Balanced Energy Regularization Loss for Out-of-Distribution Detection
Hyunjun Choi (Seoul National University), Jin Young Choi (Seoul National University)
Object DetectionSegmentationAnomaly DetectionImage
🎯 What it does: A balanced energy regularization loss is proposed to improve OOD detection performance in multi-task settings.
Balanced Product of Calibrated Experts for Long-Tailed Recognition
Emanuel Sanchez Aimar (Linköping University), Marco Kuhlmann (Linköping University)
ClassificationRecognitionMixture of ExpertsImage
🎯 What it does: A Balanced Product of Experts (BalPoE) method is proposed for long-tail classification problems, combining the weighted average of multiple logit-adjusted experts to address the head-tail category bias.
Balanced Spherical Grid for Egocentric View Synthesis
Changwoon Choi (Seoul National University), Young Min Kim (Seoul National University)
GenerationData SynthesisNeural Radiance FieldVideo
🎯 What it does: This paper proposes EgoNeRF, a neural radiance field based on balanced spherical meshes, which quickly reconstructs large-scale scenes from short-term self-captured 360° videos and achieves free-viewpoint rendering.
Balancing Logit Variation for Long-Tailed Semantic Segmentation
Yuchao Wang (Shanghai Jiao Tong University), Yujun Shen (CUHK)
SegmentationDomain AdaptationConvolutional Neural NetworkTransformerImage
🎯 What it does: The paper proposes injecting class-related logit perturbations during the training phase to alleviate the long-tail problem in semantic segmentation.
BASiS: Batch Aligned Spectral Embedding Space
Or Streicher (Technion Israel Institute of Technology), Guy Gilboa (Technion Israel Institute of Technology)
ClassificationRepresentation LearningGraph Neural NetworkImageGraph
🎯 What it does: A batch training-based direct supervised learning method called BASiS is proposed for learning the feature space of graphs (i.e., the first K eigenvectors of the graph Laplacian matrix), and the batch alignment technique is used to address the issue of inconsistent embedding coordinates across different batches.
Batch Model Consolidation: A Multi-Task Model Consolidation Framework
Iordanis Fostiropoulos (University of Southern California), Laurent Itti (University of Southern California)
ClassificationKnowledge DistillationConvolutional Neural NetworkMixture of ExpertsImageSequential
🎯 What it does: A Batch Model Consolidation (BMC) framework is proposed, which achieves long-sequence learning across multiple domains and tasks in continual learning by parallel training multiple expert models and integrating their knowledge into a baseline model through batch distillation at each step.
Bayesian Posterior Approximation With Stochastic Ensembles
Oleksandr Balabanov (Stockholm University), Hampus Linander (University of Gothenburg)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: Proposes a 'random ensemble' method that combines deep ensemble with Monte Carlo dropout/DropConnect to construct a new family of variational inference to approximate the Bayesian posterior.
BBDM: Image-to-Image Translation With Brownian Bridge Diffusion Models
Bo Li (Nanchang Hangkong University), Yu-Kun Lai (Cardiff University)
Image TranslationGenerationDiffusion modelImage
🎯 What it does: This paper proposes a diffusion model based on Brownian Bridge (BBDM) for image-to-image translation, directly modeling the random bridge diffusion process between the source domain and the target domain.
BEDLAM: A Synthetic Dataset of Bodies Exhibiting Detailed Lifelike Animated Motion
Michael J. Black (Max Planck Institute for Intelligent Systems), Jinlong Yang (Google)
Data SynthesisPose EstimationConvolutional Neural NetworkImageVideoBenchmark
🎯 What it does: This paper proposes and releases a large-scale realistic synthetic dataset called BEDLAM for training and evaluating 3D human pose and shape (HPS) estimation models.
Behavioral Analysis of Vision-and-Language Navigation Agents
Zijiao Yang (Oregon State University), Stefan Lee (Oregon State University)
TransformerVision Language ModelMultimodality
🎯 What it does: This paper studies and quantifies the fine-grained behaviors of visual-language navigation agents in four types of skills: stopping, turning, finding objects, and finding rooms, by constructing trajectory truncation and incorporating skill-specific sub-instructions as intervention samples.
Behind the Scenes: Density Fields for Single View Reconstruction
Felix Wimbauer, Daniel Cremers
Data SynthesisDepth EstimationAutonomous DrivingConvolutional Neural NetworkNeural Radiance FieldImage
🎯 What it does: This paper proposes a framework for predicting continuous density fields from a single image, which allows for voxel density evaluation at any spatial point and achieves depth prediction and novel view synthesis through volume rendering.
Being Comes From Not-Being: Open-Vocabulary Text-to-Motion Generation With Wordless Training
Junfan Lin (Sun Yat-sen University), Chang-Wen Chen
GenerationPose EstimationTransformerPrompt EngineeringVideoText
🎯 What it does: This paper proposes an offline zero-shot open-source vocabulary text-to-motion generation framework OOHMG, which utilizes prompt learning to map text to masked motions, and then reconstructs complete motions through a pre-trained Transformer motion generator.
Benchmarking Robustness of 3D Object Detection to Common Corruptions
Yinpeng Dong (Tsinghua University), Jun Zhu (Beihang University)
Object DetectionAutonomous DrivingPoint CloudBenchmark
🎯 What it does: This paper focuses on 3D object detection in autonomous driving, systematically constructing 27 common types of interference and generating corresponding corrupted versions (KITTI-C, nuScenes-C, Waymo-C) on three public datasets: KITTI, nuScenes, and Waymo, providing a benchmark for evaluating the robustness of detection models under real-world noise.
Benchmarking Self-Supervised Learning on Diverse Pathology Datasets
Mingu Kang (Lunit Inc), Sérgio Pereira (Lunit Inc)
ClassificationSegmentationConvolutional Neural NetworkTransformerContrastive LearningImageBiomedical DataBenchmark
🎯 What it does: Comparing four self-supervised learning methods on large-scale pathological images and validating their advantages in classification and nucleus segmentation tasks.
Best of Both Worlds: Multimodal Contrastive Learning With Tabular and Imaging Data
Paul Hager (Technical University of Munich), Daniel Rueckert (Technical University of Munich)
ClassificationRecognitionContrastive LearningImageMultimodalityTabularBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A multimodal framework utilizing image and tabular data for self-supervised contrastive learning is proposed, which is fine-tuned on image-only tasks, significantly improving the performance of cardiovascular disease prediction and car model recognition.
Better "CMOS" Produces Clearer Images: Learning Space-Variant Blur Estimation for Blind Image Super-Resolution
Xuhai Chen (Zhejiang University), Yong Liu (Zhejiang University)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a CMOS network for simultaneously estimating spatially varying blur kernels and semantic maps to achieve better blind super-resolution.
BEV-Guided Multi-Modality Fusion for Driving Perception
Yunze Man (University of Illinois at Urbana-Champaign), Yu-Xiong Wang (University of Illinois at Urbana-Champaign)
SegmentationAutonomous DrivingTransformerMultimodalityPoint Cloud
🎯 What it does: This paper proposes BEVGuide, an end-to-end multimodal fusion framework based on Bird’s Eye View (BEV) that can simultaneously process inputs from various sensors such as cameras, LiDAR, and millimeter-wave radar;