CVPR 2024 Papers — Page 2
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2716 papers
Adaptive VIO: Deep Visual-Inertial Odometry with Online Continual Learning
Youqi Pan, Hongbin Zha
Object DetectionDomain AdaptationConvolutional Neural NetworkSimultaneous Localization and MappingImage
🎯 What it does: This paper proposes a method for visual online adaptation aimed at improving the performance of computer vision models in dynamic environments.
AdaRevD: Adaptive Patch Exiting Reversible Decoder Pushes the Limit of Image Deblurring
Xintian Mao (Shanghai Key Laboratory of Multidimensional Information Processing East China Normal University), Yan Wang (Shanghai Key Laboratory of Multidimensional Information Processing East China Normal University)
RestorationAuto EncoderImage
🎯 What it does: For the image deblurring task, the authors propose AdaRevD, a reversible sub-decoder with an adaptive patch exit mechanism that enhances decoding capability while being memory-friendly.
AdaShift: Learning Discriminative Self-Gated Neural Feature Activation With an Adaptive Shift Factor
Sudong Cai (Kyoto University)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes the AdaShift activation function, which utilizes an adaptive shift factor to achieve more discriminative self-gated feature activation on top of ReLU.
Addressing Background Context Bias in Few-Shot Segmentation through Iterative Modulation
Lanyun Zhu (Singapore University of Technology and Design), Jun Liu (Singapore University of Technology and Design)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: This paper addresses the issue of background context bias in few-shot semantic segmentation by proposing an iterative structure-based segmentation framework. It gradually corrects the mismatch between foreground features and background information through three steps: query prediction, support modulation, and information cleaning, significantly improving segmentation performance.
ADFactory: An Effective Framework for Generalizing Optical Flow with NeRF
Han Ling (Nanjing University of Science and Technology), Xinfeng Li
Data SynthesisNeural Radiance FieldOptical FlowImageVideo
🎯 What it does: Automatically construct an Automatic Data Factory (ADF) using NeRF to generate high-quality optical flow annotations from RGB images captured by a single camera, and use them for unsupervised training of optical flow networks.
Advancing Saliency Ranking with Human Fixations: Dataset Models and Benchmarks
Bowen Deng (University of Nottingham), Michael P. Pound (University of Nottingham)
Object DetectionSegmentationGraph Neural NetworkTransformerImageBenchmark
🎯 What it does: This paper proposes a large-scale instance-level saliency ranking dataset SIFR based on real human eye gaze, and introduces a new baseline model QAGNet implemented on this dataset, aimed at accurately identifying and ranking the relative saliency of visible objects in a scene.
Adversarial Backdoor Attack by Naturalistic Data Poisoning on Trajectory Prediction in Autonomous Driving
Mozhgan Pourkeshavarz (Huawei), Amir Rasouli (Okinawa Institute of Science and Technology)
Autonomous DrivingOptimizationAdversarial AttackTime Series
🎯 What it does: A method for conducting a naturalistic backdoor attack on trajectory prediction models during the training phase has been designed.
Adversarial Distillation Based on Slack Matching and Attribution Region Alignment
Shenglin Yin (Peking University), Jieyi Long (Theta Labs)
Knowledge DistillationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a new adversarial distillation method called SmaraAD, aimed at enhancing the robustness of small models.
Adversarial Score Distillation: When score distillation meets GAN
Min Wei (Beijing University of Posts and Telecommunications), Xuesong Zhang (Beijing University of Posts and Telecommunications)
GenerationData SynthesisKnowledge DistillationDiffusion modelScore-based ModelNeural Radiance FieldGenerative Adversarial NetworkImageText
🎯 What it does: This paper proposes an Adversarial Score Distillation (ASD) method based on WGAN, treating traditional score distillation as GAN optimization to address the issues of CFG scale sensitivity and unoptimized discriminators.
Adversarial Text to Continuous Image Generation
Kilichbek Haydarov (King Abdullah University of Science and Technology), Mohamed Elhoseiny (King Abdullah University of Science and Technology)
GenerationData SynthesisGenerative Adversarial NetworkContrastive LearningImageText
🎯 What it does: This paper proposes a HyperCGAN model that generates continuous images from text descriptions.
Adversarially Robust Few-shot Learning via Parameter Co-distillation of Similarity and Class Concept Learners
Junhao Dong (Nanyang Technological University), Yew-Soon Ong (Nanyang Technological University)
Knowledge DistillationAdversarial AttackMeta LearningGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: A framework named RESISTANCE is proposed, which enhances the adversarial robustness of few-shot learning by performing parameter co-distillation between two branches of visual similarity learning and category concept learning in adversarial training.
Aerial Lifting: Neural Urban Semantic and Building Instance Lifting from Aerial Imagery
Yuqi Zhang (Sun Yat-sen University), Shuguang Cui (Sun Yat-sen University)
Object DetectionSegmentationNeural Radiance FieldContrastive LearningImage
🎯 What it does: Utilizing Neural Radiance Fields (NeRF) to elevate 2D semantic and instance segmentation labels from a drone's perspective to 3D space, achieving semantic and building-level instance segmentation at an urban scale.
AEROBLADE: Training-Free Detection of Latent Diffusion Images Using Autoencoder Reconstruction Error
Jonas Ricker (Ruhr University Bochum), Asja Fischer (Ruhr University Bochum)
GenerationAnomaly DetectionDiffusion modelAuto EncoderImage
🎯 What it does: A training-free detection method called AEROBLADE is proposed, which utilizes the reconstruction error of a latent diffusion model (LDM) autoencoder to distinguish between real images and generated images.
AETTA: Label-Free Accuracy Estimation for Test-Time Adaptation
Taeckyung Lee (KAIST), Sung-Ju Lee (KAIST)
Domain AdaptationImage
🎯 What it does: A test-time adaptive accuracy estimation method called AETTA is proposed, which does not require labels or source data.
Affine Equivariant Networks Based on Differential Invariants
Yikang Li (Peking University), Zhouchen Lin (Peking University)
ImageOrdinary Differential Equation
🎯 What it does: An affine equivariant network based on differential invariants and symmetric partial differential equations (InvarPDEs-Net and InvarLayer) is constructed, which does not require discretization or sampling of the affine group.
AHIVE: Anatomy-aware Hierarchical Vision Encoding for Interactive Radiology Report Retrieval
Sixing Yan (Hong Kong Baptist University), Simon See (NVIDIA Corporation)
RetrievalTransformerContrastive LearningImageTextMultimodalityElectronic Health Records
🎯 What it does: A medical image retrieval framework based on Hierarchical Anatomical Visual Encoding (AHIVE) is proposed, which can retrieve radiology reports that are highly relevant and clinically accurate from X-ray images, and supports interactive revision.
AIDE: An Automatic Data Engine for Object Detection in Autonomous Driving
Mingfu Liang (Northwestern University), Manmohan Chandraker (NEC Laboratories America)
Object DetectionAutonomous DrivingLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: An automatic data engine AIDE is proposed for automatically discovering, collecting, pseudo-labeling, and validating novel object detection in autonomous driving scenarios, reducing the cost of manual annotation.
AiOS: All-in-One-Stage Expressive Human Pose and Shape Estimation
Qingping Sun (SenseTime Research), Zhongang Cai (SenseTime Research)
Pose EstimationTransformerImage
🎯 What it does: The first end-to-end integrated stage (AiOS) framework is proposed, achieving expression human pose and shape (EHPS) recovery without additional detection networks, directly predicting the SMPL-X 3D mesh for the whole body, hands, and face from the entire image.
AirPlanes: Accurate Plane Estimation via 3D-Consistent Embeddings
Jamie Watson (Niantic), Sara Vicente (Niantic)
SegmentationPoint Cloud
🎯 What it does: A 3D plane embedding learning and clustering framework based on multi-view RGB images is proposed for high-precision plane segmentation.
Alchemist: Parametric Control of Material Properties with Diffusion Models
Prafull Sharma (Google Research), Mark Matthews (Massachusetts Institute of Technology)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: By training an implicit diffusion model based on Stable Diffusion, we achieve parametric and continuous editing of material properties such as roughness, metallicity, diffuse reflection, and transparency of objects in images.
ALGM: Adaptive Local-then-Global Token Merging for Efficient Semantic Segmentation with Plain Vision Transformers
Narges Norouzi (Eindhoven University of Technology), Gijs Dubbelman (Eindhoven University of Technology)
SegmentationComputational EfficiencyTransformerImage
🎯 What it does: An adaptive local-global token merging method (ALGM) is proposed, which can significantly reduce the number of tokens in ViT while maintaining semantic segmentation accuracy and improving inference speed.
Align and Aggregate: Compositional Reasoning with Video Alignment and Answer Aggregation for Video Question-Answering
Zhaohe Liao (Shanghai Jiao Tong University), Liqing Zhang (Shanghai Jiao Tong University)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningVideoMultimodality
🎯 What it does: A model-agnostic Video Alignment and Answer Aggregation (VA³) framework is proposed to enhance the interpretability and logical consistency of video question answering (VidQA) through hierarchical video alignment and answer aggregation.
Align Before Adapt: Leveraging Entity-to-Region Alignments for Generalizable Video Action Recognition
Yifei Chen (Huawei Technologies), Wei Peng (Huawei Technologies)
RecognitionDomain AdaptationComputational EfficiencyTransformerVision Language ModelVideo
🎯 What it does: This paper proposes an 'Align-before-Adapt' paradigm, using entity-region alignment to guide the training of video action recognition models.
Align Your Gaussians: Text-to-4D with Dynamic 3D Gaussians and Composed Diffusion Models
Huan Ling (NVIDIA), Karsten Kreis (NVIDIA)
GenerationData SynthesisDiffusion modelScore-based ModelVideoText
🎯 What it does: A text-driven 4D scene generation method based on dynamic 3D Gaussian and composite diffusion models is proposed, which can generate animatable 3D objects from text and combine them into large dynamic scenes.
Aligning and Prompting Everything All at Once for Universal Visual Perception
Yunhang Shen (Tencent), Rongrong Ji (Xiamen University)
Object DetectionSegmentationTransformerPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: A unified visual perception model APE has been constructed, capable of detecting, segmenting, and localizing all targets, segmented areas, and natural language descriptions in an image at once, using an instance-level sentence-object matching framework.
Aligning Logits Generatively for Principled Black-Box Knowledge Distillation
Jing Ma (Huazhong University of Science and Technology), Yongbin Li (Alibaba Damo Academy)
Knowledge DistillationGenerative Adversarial NetworkImage
🎯 What it does: This paper studies the black-box knowledge distillation (B2KD) problem from cloud to edge, proposing a two-step process: first, deprivatization is achieved through a generator, and then distillation is performed using mapping emulation.
AlignMiF: Geometry-Aligned Multimodal Implicit Field for LiDAR-Camera Joint Synthesis
Tang Tao (Shenzhen Campus of Sun Yat-sen University), Xiaodan Liang (DarkMatter AI Research)
Data SynthesisAutonomous DrivingNeural Radiance FieldMultimodalityPoint Cloud
🎯 What it does: A multi-modal implicit field AlignMiF is proposed for the joint synthesis of LiDAR and camera data.
AlignSAM: Aligning Segment Anything Model to Open Context via Reinforcement Learning
Duojun Huang (Sun Yat-sen University), Guanbin Li (Sun Yat-sen University)
Object DetectionSegmentationReinforcement LearningPrompt EngineeringImage
🎯 What it does: Proposes AlignSAM, a reinforcement learning-based automatic prompting framework that aligns the Segment Anything Model to different open-source task scenarios without changing SAM parameters;
All in One Framework for Multimodal Re-identification in the Wild
He Li (Wuhan University), Bo Du
RecognitionRetrievalTransformerContrastive LearningImageTextMultimodality
🎯 What it does: A unified framework AIO is proposed, using a frozen large foundation model and a multimodal tokenizer to map four modalities: RGB, IR, Sketch, and Text to a shared embedding space, achieving zero-shot cross-modal and multimodal portrait recognition.
All Rivers Run to the Sea: Private Learning with Asymmetric Flows
Yue Niu (University of Southern California), Salman Avestimehr (University of Southern California)
Federated LearningSafty and PrivacyComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: The Delta framework is designed and implemented, which splits the intermediate representations of deep networks into information-sensitive low-dimensional parts and residual high-dimensional parts, trained and inferred in a constrained private environment and a high-performance public GPU, respectively.
AllSpark: Reborn Labeled Features from Unlabeled in Transformer for Semi-Supervised Semantic Segmentation
Haonan Wang (Hong Kong University of Science and Technology), Xiaomeng Li (Hong Kong University of Science and Technology)
SegmentationTransformerImage
🎯 What it does: The AllSpark module is proposed to reshape labeled features by mapping unlabeled features to labeled features in semi-supervised semantic segmentation, alleviating the issue of label data dominance.
Alpha Invariance: On Inverse Scaling Between Distance and Volume Density in Neural Radiance Fields
Joshua Ahn (University of Chicago), Greg Shakhnarovich (TTI-Chicago)
Neural Radiance FieldPoint Cloud
🎯 What it does: This paper addresses the inverse relationship between volume density and distance caused by the uncertainty of scene size scale in NeRF, proposing the concept of alpha invariance and providing corresponding mathematical analysis.
Alpha-CLIP: A CLIP Model Focusing on Wherever You Want
Zeyi Sun (Shanghai Jiao Tong University), Jiaqi Wang (Shanghai AI Laboratory)
ClassificationObject DetectionSegmentationGenerationTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: An alpha channel is added to the CLIP model, enabling it to focus on specified areas through masking, resulting in Alpha-CLIP obtained by fine-tuning the original CLIP weights.
AM-RADIO: Agglomerative Vision Foundation Model Reduce All Domains Into One
Mike Ranzinger (NVIDIA), Pavlo Molchanov (NVIDIA)
RecognitionObject DetectionSegmentationKnowledge DistillationTransformerContrastive LearningImage
🎯 What it does: By fusing visual foundation models such as CLIP, DINOv2, and SAM through multi-teacher distillation, a RADIO visual model is trained that can handle arbitrary resolutions and is compatible with multiple tasks.
Amodal Completion via Progressive Mixed Context Diffusion
Katherine Xu (University of Pennsylvania), Jianbo Shi (University of Pennsylvania)
RestorationGenerationDiffusion modelImage
🎯 What it does: A training-free progressive occlusion-aware generation method is proposed to restore unobstructed pixels of occluded objects in natural images.
Amodal Ground Truth and Completion in the Wild
Guanqi Zhan (VGG University of Oxford), Andrew Zisserman (VGG University of Oxford)
Object DetectionSegmentationConvolutional Neural NetworkDiffusion modelImageMesh
🎯 What it does: This paper automatically generates complete (amodal) segmentation masks for partially occluded objects in real images using 3D scene data, and based on this, constructs a large-scale multi-category MP3D-Amodal dataset. It also proposes two amodal completion models that do not require manually provided occluder masks: a two-stage OccAmodal and a single-stage SDAmodal based on Stable Diffusion.
AMU-Tuning: Effective Logit Bias for CLIP-based Few-shot Learning
Yuwei Tang (Tianjin University), Qinghua Hu (Tianjin University)
ClassificationRecognitionTransformerContrastive LearningImage
🎯 What it does: The AMU-Tuning method is proposed to enhance the performance of CLIP in few-shot learning by learning effective logit biases, and it provides a unified analysis of existing methods from the perspective of logit biases.
An Aggregation-Free Federated Learning for Tackling Data Heterogeneity
Yuan Wang (Institute of High Performance Computing Agency for Science Technology and Research), Rick Siow Mong Goh (Institute of High Performance Computing Agency for Science Technology and Research)
Federated LearningImage
🎯 What it does: This paper proposes FedAF, which adopts an aggregation-independent federated learning framework to address the client drift problem caused by data heterogeneity.
An Asymmetric Augmented Self-Supervised Learning Method for Unsupervised Fine-Grained Image Hashing
Feiran Hu (Nanjing University of Science and Technology), Lingyan Gao (AInnovation Technology Group Co., Ltd)
RetrievalAuto EncoderContrastive LearningImage
🎯 What it does: This paper proposes an unsupervised fine-grained image hashing method A2-SSL, which utilizes asymmetric augmentation, self-contrast learning, and self-consistent reconstruction to generate high-quality binary codes for fine-grained image retrieval.
An Edit Friendly DDPM Noise Space: Inversion and Manipulations
Inbar Huberman-Spiegelglas (Technion Israel Institute of Technology), Tomer Michaeli (Technion Israel Institute of Technology)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: This paper proposes an edit-friendly noise space for DDPM and provides a fast, optimization-free inversion method that can extract a set of noise vectors from any real image, which are easy to edit while preserving the image structure.
An Empirical Study of Scaling Law for Scene Text Recognition
Miao Rang (Huawei Noah's Ark Lab), Kai Han (Huawei Noah's Ark Lab)
RecognitionTransformerImageText
🎯 What it does: This study investigates the scaling relationship between model size, data volume, and computational resources in scene text recognition (STR), constructing a large-scale mixed dataset REBU-Syn, and training and validating the scaling law based on Transformer.
An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains
George Eskandar (University of Stuttgart)
Object DetectionDomain AdaptationAutonomous DrivingTransformerPoint CloudBenchmark
🎯 What it does: A systematic empirical study on the generalization ability of LiDAR-based 3D object detectors under different domains (sensors, weather, geographical locations).
An Interactive Navigation Method with Effect-oriented Affordance
Xiaohan Wang (Xi'an Jiaotong University), Shuqiang Jiang (Chinese Academy of Sciences)
Object DetectionOptimizationRobotic IntelligenceConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: This paper proposes an interactive navigation method based on effect-driven affordance maps (ADIN), achieving long-term goal navigation in complex dynamic environments.
An N-Point Linear Solver for Line and Motion Estimation with Event Cameras
Ling Gao (ShanghaiTech University), Laurent Kneip (University of Zurich)
Pose EstimationOptimizationSimultaneous Localization and MappingTime SeriesBenchmark
🎯 What it does: A linear multi-point solver based on event cameras is proposed, capable of estimating the camera's linear velocity and line parameters from an event stream in one go.
An Upload-Efficient Scheme for Transferring Knowledge From a Server-Side Pre-trained Generator to Clients in Heterogeneous Federated Learning
Jianqing Zhang (Shanghai Jiao Tong University), Jian Cao (Shanghai Jiao Tong University)
Federated LearningKnowledge DistillationDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a high-efficiency knowledge transfer scheme called FedKTL, which uses a server-side pre-trained generator to generate image-vector pairs related to client tasks, aiding the transfer of global knowledge to client models in heterogeneous federated learning.
Analyzing and Improving the Training Dynamics of Diffusion Models
Tero Karras (NVIDIA), Samuli Laine (NVIDIA)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Improved the training dynamics of the diffusion model, redesigned network layers to maintain the expected values of activations, weights, and update magnitudes, and proposed a post-EMA method to accurately adjust the model average;
Anatomically Constrained Implicit Face Models
Prashanth Chandran (Disney Research Studios), Gaspard Zoss (Disney Research Studios)
GenerationOptimizationMesh
🎯 What it does: A novel facial model called the Anatomical Implicit Face Model (AIM) has been constructed, which learns both the facial skeletal structure and the epidermal skin surface simultaneously through implicit neural networks. This model can achieve high-precision actor-specific Blendshape models and supports efficient model fitting, editing, and performance redirection.
Anchor-based Robust Finetuning of Vision-Language Models
Jinwei Han (Wuhan University), Gui-Song Xia (Wuhan University)
RetrievalDomain AdaptationTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposes Anchor-based Robust Finetuning (ARF), which introduces two auxiliary anchors (text compensation anchor and image-text retrieval anchor) during the fine-tuning of CLIP for downstream tasks to maintain the original model's OOD generalization ability.
ANIM: Accurate Neural Implicit Model for Human Reconstruction from a single RGB-D Image
Marco Pesavento (University of Surrey), Tony Tung (Meta Reality Labs)
SegmentationPose EstimationDepth EstimationConvolutional Neural NetworkNeural Radiance FieldImageMultimodality
🎯 What it does: ANIM is constructed, a single-view RGB-D human reconstruction method based on a neural implicit model.
Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling
Zhe Li (Tsinghua University), Yebin Liu (Tsinghua University)
GenerationData SynthesisPose EstimationGenerative Adversarial NetworkGaussian SplattingVideo
🎯 What it does: Using multi-view RGB video to learn animatable human avatars with high-fidelity details, generating pose-related Gaussian maps and rendering them into images through 3D Gaussian splatting.
Animate Anyone: Consistent and Controllable Image-to-Video Synthesis for Character Animation
Li Hu (Alibaba Group)
GenerationData SynthesisPose EstimationDiffusion modelVideo
🎯 What it does: A full-process method based on diffusion models is proposed, which synthesizes continuous, high-quality videos from a single character image according to a specified pose sequence.
Animating General Image with Large Visual Motion Model
Dengsheng Chen (Meituan), Xiaolin Wei (Meituan)
GenerationData SynthesisDiffusion modelImageVideo
🎯 What it does: Proposes the Large Visual Motion Model (LVMM), which can predict local motion trajectories from a single image and generate high-quality dynamic videos.
Anomaly Heterogeneity Learning for Open-set Supervised Anomaly Detection
Jiawen Zhu (Singapore Management University), Guansong Pang (Harvard University)
Anomaly DetectionRecurrent Neural NetworkImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A framework for open-set supervised anomaly detection based on simulated heterogeneous anomaly distributions (AHL) is proposed, which achieves unified modeling of seen and unseen anomalies by generating diverse anomaly distributions and performing collaborative differentiable learning in a simulated open-set environment.
Anomaly Score: Evaluating Generative Models and Individual Generated Images based on Complexity and Vulnerability
Jaehui Hwang (Yonsei University), Jong-Seok Lee (Yonsei University)
GenerationAnomaly DetectionConvolutional Neural NetworkTransformerDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: Two new metrics based on the complexity and vulnerability of image representation space are proposed to measure the naturalness of generated images. Anomaly scores (AS) for evaluating generative models and anomaly scores for individual generated images (AS-i) are constructed based on these two metrics.
Any-Shift Prompting for Generalization over Distributions
Zehao Xiao (University of Amsterdam), Cees G. M. Snoek (University of Amsterdam)
Domain AdaptationTransformerPrompt EngineeringVision Language ModelImageText
🎯 What it does: In the prompt learning of the visual-language model CLIP, a framework is proposed that can automatically generate test-specific prompts for any distribution shift.
AnyDoor: Zero-shot Object-level Image Customization
Xi Chen (University of Hong Kong), Hengshuang Zhao (University of Hong Kong)
Image TranslationGenerationDiffusion modelImageVideo
🎯 What it does: A zero-shot object-level image customization framework called AnyDoor is proposed, which can transfer target objects to any scene at specified locations and shapes without parameters.
AnyScene: Customized Image Synthesis with Composited Foreground
Ruidong Chen (Tianjin University), An-An Liu (Tianjin University)
Image HarmonizationGenerationData SynthesisDiffusion modelAuto EncoderImageTextBenchmark
🎯 What it does: The AnyScene framework is proposed, which synthesizes diverse and visually harmonious background scenes based on user-provided foreground images and text prompts.
AnySkill: Learning Open-Vocabulary Physical Skill for Interactive Agents
Jieming Cui (Peking University), Siyuan Huang (National Key Laboratory of General Artificial Intelligence)
Robotic IntelligenceTransformerReinforcement LearningVision Language ModelImageText
🎯 What it does: We designed and implemented AnySkill, a hierarchical framework for learning physical skills. It first learns physically feasible atomic actions using low-level controllers, and then combines these atomic actions through high-level strategies under open vocabulary text instructions using CLIP image similarity rewards to generate natural and interactive human movements.
APISR: Anime Production Inspired Real-World Anime Super-Resolution
Boyang Wang (University of Michigan), Hanbin Zhao (Zhejiang University)
RestorationSuper ResolutionCompressionGenerative Adversarial NetworkImageVideo
🎯 What it does: This paper proposes an image dataset API based on anime production by analyzing the anime production process, and designs a prediction-guided compression module and pseudo-realistic line enhancement technology to address compression distortion and hand-drawn line blur.
APSeg: Auto-Prompt Network for Cross-Domain Few-Shot Semantic Segmentation
Weizhao He (Shenzhen University), Liang Sun (Shenzhen University)
SegmentationDomain AdaptationMeta LearningTransformerPrompt EngineeringImageMagnetic Resonance Imaging
🎯 What it does: This paper proposes a fully automatic cross-domain few-shot semantic segmentation framework called APSeg, which achieves rapid adaptation to the target domain through dual-prototype anchor point transformation and meta-prompt generation.
Arbitrary Motion Style Transfer with Multi-condition Motion Latent Diffusion Model
Wenfeng Song (Beijing Information Science and Technology University), Hong Qin (Stony Brook University)
GenerationData SynthesisTransformerDiffusion model
🎯 What it does: Proposes a Multi-Conditional Motion Latent Diffusion Model (MCM-LDM) for arbitrary motion style transfer.
Arbitrary-Scale Image Generation and Upsampling using Latent Diffusion Model and Implicit Neural Decoder
Jinseok Kim (KAIST), Tae-Kyun Kim (Imperial College London)
GenerationSuper ResolutionDiffusion modelAuto EncoderImage
🎯 What it does: A pipeline that combines the Latent Diffusion Model (LDM) and an implicit neural decoder (based on LIIF) is proposed to achieve image generation and super-resolution at arbitrary scales;
Are Conventional SNNs Really Efficient? A Perspective from Network Quantization
Guobin Shen (Chinese Academy of Sciences Institute of Automation), Yi Zeng (Chinese Academy of Sciences Institute of Automation)
Computational EfficiencySpiking Neural NetworkImage
🎯 What it does: This paper conducts a unified quantitative analysis of the energy consumption and computational efficiency of SNNs and quantized ANNs, proposing a 'Bit Budget' framework to allocate bits for weight, activation quantization, and time steps, thereby optimizing SNN design.
ArGue: Attribute-Guided Prompt Tuning for Vision-Language Models
Xinyu Tian (Australian National University), Jing Zhang (Australian National University)
ClassificationRecognitionDomain AdaptationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: This study investigates the use of visual attribute-guided prompt tuning to enhance the generalization and robustness of visual language models, proposing two methods: ArGue and ArGue-N.
ArtAdapter: Text-to-Image Style Transfer using Multi-Level Style Encoder and Explicit Adaptation
Dar-Yen Chen (PicCollage), Ching-Wen Hsu (PicCollage)
Image TranslationGenerationDiffusion modelImageText
🎯 What it does: Proposes ArtAdapter, a text-to-image style transfer framework that utilizes a multi-layer style encoder and an explicit adaptation mechanism;
Artist-Friendly Relightable and Animatable Neural Heads
Yingyan Xu (ETH Zurich), Derek Bradley (Disney Research Studios)
GenerationData SynthesisConvolutional Neural NetworkMixture of ExpertsVideoMesh
🎯 What it does: A neural head model that can be rendered under any lighting and expression conditions is proposed, combining MVP with local lighting/viewpoint encoding to achieve dynamic relighting and animation control.
ARTrackV2: Prompting Autoregressive Tracker Where to Look and How to Describe
Yifan Bai (Xi'an Jiaotong University), Xing Wei (Xi'an Jiaotong University)
Object TrackingTransformerVideo
🎯 What it does: ARTrackV2 is an autoregressive visual tracker based on a generative framework, capable of simultaneously predicting the trajectory position and appearance features of the target.
As-Plausible-As-Possible: Plausibility-Aware Mesh Deformation Using 2D Diffusion Priors
Seungwoo Yoo, Minhyuk Sung
GenerationOptimizationDiffusion modelScore-based ModelMeshBenchmark
🎯 What it does: This paper proposes a perceptual mesh deformation method called APAP, which allows users to edit meshes while preserving details by manipulating control points.
ASAM: Boosting Segment Anything Model with Adversarial Tuning
Bo Li (Vivo Mobile Communication Co., Ltd), Lv Tang (Vivo Mobile Communication Co., Ltd)
SegmentationAdversarial AttackDiffusion modelImage
🎯 What it does: Adversarial fine-tuning of the Segment Anything Model (SAM) is performed using a stable diffusion model to generate natural adversarial samples to enhance the model's segmentation performance.
ASH: Animatable Gaussian Splats for Efficient and Photoreal Human Rendering
Haokai Pang (Max Planck Institute for Informatics), Marc Habermann (Max Planck Institute for Informatics)
Image TranslationPose EstimationComputational EfficiencyGaussian SplattingVideoMesh
🎯 What it does: This paper proposes a real-time rendering method for animated high-fidelity human portraits called ASh, which attaches 3D Gaussian point clouds to a deformable template mesh and learns its parameters in texture space, enabling controllable rendering based on skeletal poses.
AssistGUI: Task-Oriented PC Graphical User Interface Automation
Difei Gao, Mike Zheng Shou
TransformerLarge Language ModelAgentic AITextMultimodalityBenchmark
🎯 What it does: Proposed the ASSISTGUI benchmark, covering 100 Windows PC GUI automation tasks, and introduced the AutoPC multi-agent collaboration framework to accomplish these tasks;
Asymmetric Masked Distillation for Pre-Training Small Foundation Models
Zhiyu Zhao (Nanjing University), Limin Wang (Nanjing University)
Knowledge DistillationRepresentation LearningTransformerAuto EncoderImageVideo
🎯 What it does: An Asynchronous Mask Distillation framework (AMD) is proposed for self-supervised pre-training on small Vision Transformers.
Atlantis: Enabling Underwater Depth Estimation with Stable Diffusion
Fan Zhang (Beijing Institute of Technology), Ying Fu (Beijing Institute of Technology)
GenerationData SynthesisDepth EstimationDiffusion modelImageText
🎯 What it does: Using Stable Diffusion combined with ControlNet to generate realistic underwater images and corresponding accurate depths, and based on this dataset to train and enhance the performance of monocular depth estimation models.
Atom-Level Optical Chemical Structure Recognition with Limited Supervision
Martijn Oldenhof (KU Leuven), Yves Moreau (KU Leuven)
RecognitionObject DetectionImageGraph
🎯 What it does: A weakly supervised learning-based chemical structure recognition model, AtomLenz, has been developed, capable of achieving atomic-level localization and constructing molecular graphs using only SMILES-labeled image data.
Attack To Defend: Exploiting Adversarial Attacks for Detecting Poisoned Models
Samar Fares (Mohamed bin Zayed University of Artificial Intelligence), Karthik Nandakumar (Mohamed bin Zayed University of Artificial Intelligence)
Anomaly DetectionAdversarial AttackImage
🎯 What it does: This paper proposes a framework A2D (Attack To Defend) for detecting whether a model has been implanted with backdoors based on adversarial attack detection.
Attention Calibration for Disentangled Text-to-Image Personalization
Yanbing Zhang (East China University of Science and Technology), Zhe Wang (East China University of Science and Technology)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: DisenDiff is designed to learn and generate various personalized concepts in a text-to-image model using an attention calibration mechanism under the condition of having only a single image, supporting both independent and combined generation of concepts.
Attention-Driven Training-Free Efficiency Enhancement of Diffusion Models
Hongjie Wang (Princeton University), Yuchen Liu (Adobe Research)
GenerationComputational EfficiencyDiffusion modelImage
🎯 What it does: Proposes the AT-EDM framework, which utilizes attention maps for dynamic token pruning of diffusion models (mainly SD-XL) during inference without training, thereby accelerating the generation process.
Attention-Propagation Network for Egocentric Heatmap to 3D Pose Lifting
Taeho Kang (Seoul National University), Youngki Lee (Seoul National University)
Pose EstimationTransformerImage
🎯 What it does: A heatmap-based 3D pose upsampling method called EgoTAP is proposed for high-precision 3D human pose estimation in a binocular first-person camera scenario.
Attentive Illumination Decomposition Model for Multi-Illuminant White Balancing
Dongyoung Kim (Yonsei University), Seon Joo Kim (Yonsei University)
Image TranslationRestorationConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: An end-to-end multi-illumination white balance model AID is proposed, which utilizes slot attention to decompose the chromaticity and pixel-level weights of each light source in the scene, and synthesizes the final illumination map based on a linear mixing model.
Attribute-Guided Pedestrian Retrieval: Bridging Person Re-ID with Internal Attribute Variability
Yan Huang (Chinese Academy of Sciences), Liang Wang (Chinese Academy of Sciences)
RetrievalTransformerImageBenchmark
🎯 What it does: A Transformer-based Attribute-Guided Pedestrian Retrieval framework (ATPR) is proposed to achieve retrieval of the same identity under different attribute conditions.
AttriHuman-3D: Editable 3D Human Avatar Generation with Attribute Decomposition and Indexing
Fan Yang (Nanyang Technological University), Guosheng Lin (Nanyang Technological University)
GenerationGenerative Adversarial NetworkImage
🎯 What it does: This paper presents AttrIHuman-3D, an editable 3D human head generation model that supports attribute-based fine-grained editing.
Audio-Visual Segmentation via Unlabeled Frame Exploitation
Jinxiang Liu (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)
SegmentationOptical FlowVideo
🎯 What it does: This paper proposes a unified framework for utilizing unlabeled frames, categorizing them into adjacent frames and remote frames based on temporal features in the audio-video segmentation task, and employing optical flow motion guidance and a teacher-student self-supervised enhancement model for each category;
AUEditNet: Dual-Branch Facial Action Unit Intensity Manipulation with Implicit Disentanglement
Shiwei Jin (University of California San Diego), Truong Nguyen (University of California San Diego)
Image TranslationGenerationGenerative Adversarial NetworkImage
🎯 What it does: AUEditNet is proposed, which can achieve fine editing of the intensity of 12 facial action units (AUs) in high-resolution facial images based on data from only 18 subjects, and supports conditional control based on intensity values or target images.
Authentic Hand Avatar from a Phone Scan via Universal Hand Model
Gyeongsik Moon (Meta), Takaaki Shiratori (Meta)
GenerationPose EstimationConvolutional Neural NetworkAuto EncoderOptical FlowMesh
🎯 What it does: A Universal Hand Model (UHM) is proposed, and a pipeline is constructed to quickly generate realistic 3D hand avatars from mobile scans, capable of simultaneously performing hand pose tracking and shape modeling, and adapting to mobile scan data to produce animatable high-fidelity hand meshes and textures.
Auto MC-Reward: Automated Dense Reward Design with Large Language Models for Minecraft
Hao Li (CUHK-SenseTime Joint Laboratory, Chinese University of Hong Kong), Jifeng Dai (Tsinghua University)
TransformerLarge Language ModelReinforcement LearningVideo
🎯 What it does: This paper presents Auto MC-Reward, a framework for automatically generating and iteratively refining dense reward functions for sparse reward tasks in the Minecraft environment based on large language models.
Auto-Train-Once: Controller Network Guided Automatic Network Pruning from Scratch
Xidong Wu (University of Pittsburgh), Heng Huang
CompressionOptimizationConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: Proposes the Auto-Train-Once (ATO) framework, an end-to-end structural pruning method that achieves one-time training and pruning.
AutoAD III: The Prequel - Back to the Pixels
Tengda Han (University of Oxford), Andrew Zisserman (University of Oxford)
GenerationTransformerLarge Language ModelVision Language ModelVideoMultimodalityAudio
🎯 What it does: This paper proposes two pixel-level movie audio description (AD) datasets (CMD-AD and HowTo-AD), and based on this, designs the Movie-BLIP2 and Movie-Llama2 models based on Q-former to achieve full pixel-level AD generation.
Automatic Controllable Colorization via Imagination
Xiaoyan Cong (Zhejiang University), Chenyang Lei
SegmentationGenerationTransformerDiffusion modelImage
🎯 What it does: An automatic and controllable image colorization framework is proposed, utilizing a pre-trained diffusion model to generate diverse reference images and achieving colorization through paragraph-level selection.
Autoregressive Queries for Adaptive Tracking with Spatio-Temporal Transformers
Jinxia Xie (Guangxi Normal University), Rongrong Ji (Xiamen University)
Object TrackingTransformerVideo
🎯 What it does: A spatiotemporal Transformer tracker called AQATrack is proposed, which can learn the appearance changes of the target across consecutive frames using autoregressive queries.
AV-RIR: Audio-Visual Room Impulse Response Estimation
Anton Ratnarajah (University of Maryland), Dinesh Manocha (University of Maryland)
RecognitionRestorationConvolutional Neural NetworkContrastive LearningMultimodalityAudio
🎯 What it does: By utilizing joint learning of audio and visual signals, the AV-RIR model is proposed to achieve high-precision estimation of room impulse response (RIR), with dereverberation as an auxiliary task.
AV2AV: Direct Audio-Visual Speech to Audio-Visual Speech Translation with Unified Audio-Visual Speech Representation
Jeongsoo Choi (Korea Advanced Institute of Science and Technology), Yong Man Ro (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisTransformerGenerative Adversarial NetworkVideoMultimodalityAudio
🎯 What it does: A framework called AV2AV is proposed, which directly translates audio-video speech to audio-video speech.
AvatarGPT: All-in-One Framework for Motion Understanding Planning Generation and Beyond
Zixiang Zhou, Baoyuan Wang
GenerationRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningVideoText
🎯 What it does: The AvatarGPT framework is proposed, integrating seven tasks including action understanding, planning, decomposition, generation, and intermediate generation into a closed loop, creating a unified generation system from natural language to long sequences of human actions.
AVFF: Audio-Visual Feature Fusion for Video Deepfake Detection
Trevine Oorloff (University of Maryland), Gaurav Bharaj (Reality Defender Inc)
ClassificationRecognitionTransformerAuto EncoderContrastive LearningVideoMultimodalityAudio
🎯 What it does: A two-stage audio-visual fusion model (AVFF) is proposed, which first learns the audio-visual correspondence through self-supervised learning on real videos, and then performs supervised classification on deepfake data to achieve video deepfake detection.
AVID: Any-Length Video Inpainting with Diffusion Model
Zhixing Zhang (Rutgers University), Licheng Yu (Meta)
RestorationGenerationDiffusion modelVideo
🎯 What it does: This paper proposes a method for arbitrary-length video inpainting based on diffusion models, called AVID, which can generate semantically consistent content that matches the original video in any masked area based on text prompts.
AZ-NAS: Assembling Zero-Cost Proxies for Network Architecture Search
Junghyup Lee (Yonsei University), Bumsub Ham (Korea Institute of Science and Technology)
Neural Architecture SearchTransformerImage
🎯 What it does: A training-free neural architecture search method called AZ-NAS is proposed, which evaluates network performance using a combination of various zero-cost proxies and outputs rankings.
BA-SAM: Scalable Bias-Mode Attention Mask for Segment Anything Model
Yiran Song (Shanghai Jiao Tong University), Lizhuang Ma (La Trobe University)
SegmentationTransformerImage
🎯 What it does: This paper addresses the performance degradation of the Segment Anything Model at different image resolutions by proposing a scalable Bias Attention Mask (BA-SAM), achieving length extrapolation without altering the structure.
Back to 3D: Few-Shot 3D Keypoint Detection with Back-Projected 2D Features
Thomas Wimmer (École Polytechnique), Maks Ovsjanikov (Technical University of Munich)
Object DetectionSegmentationOptimizationTransformerContrastive LearningPoint CloudMesh
🎯 What it does: Using features from a large-scale pre-trained 2D vision model, 3D features with semantic and geometric information are constructed through multi-view rendering and back-projection onto a 3D mesh; based on the similarity of these features to keypoint candidates and geometric distances, an optimization module is used to achieve few-shot keypoint detection.
Backdoor Defense via Test-Time Detecting and Repairing
Jiyang Guan (Chinese Academy of Sciences), Ran He (Chinese Academy of Sciences)
Anomaly DetectionAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A two-stage post-training reverse gating defense framework based on test-time data, TTBD, is proposed for detecting and removing backdoor attacks in deep models.
Backpropagation-free Network for 3D Test-time Adaptation
Yanshuo Wang (Australian National University), Mehrtash Harandi (Monash University)
Domain AdaptationPoint Cloud
🎯 What it does: A gradient-free backpropagation method for 3D point cloud adaptation during testing, BFTT3D, is proposed. It utilizes a non-parametric network to extract features, achieves alignment between the source and target domains through prototype memory and subspace learning, and employs entropy adaptive fusion to generate the final predictions.
BadCLIP: Dual-Embedding Guided Backdoor Attack on Multimodal Contrastive Learning
Siyuan Liang (National University of Singapore), Ee-Chien Chang (National University of Singapore)
Representation LearningAdversarial AttackTransformerContrastive LearningMultimodality
🎯 What it does: A dual-embedding guided backdoor attack framework BadCLIP is proposed for the multimodal contrastive learning model CLIP;