IEEE/CVF Conference on Computer Vision and Pattern Recognition · 851 papers
Practical Solutions to the Relative Pose of Three Calibrated Cameras
Charalambos Tzamos (Czech Technical University in Prague), Zuzana Kukelova (Czech Technical University in Prague)
CodePose EstimationSimultaneous Localization and MappingImage
🎯 What it does: For the problem of estimating the relative pose of three calibrated cameras using four points and three views, two types of approximate geometric-based minimal solvers (4p3v(A) and 4p3v(M)) are proposed, which are further improved within the RANSAC framework to include approximate corresponding point offsets, early non-minimal re-estimation, three-view filtering, and local optimization to form a complete solver.
🎯 What it does: A training-agnostic concept elimination method called AdaVD has been developed, which can accurately erase target concepts in text-to-image diffusion models while preserving prior knowledge of non-target concepts.
Preconditioners for the Stochastic Training of Neural Fields
Shin-Fang Chng (Australian Institute of Machine Learning, University of Adelaide), Simon Lucey (Australian Institute of Machine Learning, University of Adelaide)
CodeOptimizationNeural Radiance FieldImage
🎯 What it does: A curvature-aware preconditioning framework for the random training of neural fields is proposed, and its acceleration effects are validated under various activation functions.
Xiuqiang Song (Shandong University), Xueying Qin (Qilu University of Technology)
CodeObject TrackingPose EstimationSimultaneous Localization and MappingImagePoint CloudMesh
🎯 What it does: This paper proposes a 3D object tracking method called BIT, which completely relies on RGB images to generate high-precision mesh models in real-time and accomplish 6DoF pose tracking without any pre-trained models or training data.
Probabilistic Prompt Distribution Learning for Animal Pose Estimation
Jiyong Rao (Tongji University), Yu Wang (Tongji University)
CodePose EstimationTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: This paper proposes an animal pose estimation method based on probabilistic prompt distribution learning, utilizing learnable prompts and cross-modal fusion to enhance cross-species generalization capabilities.
Progressive Focused Transformer for Single Image Super-Resolution
Wei Long (University of Electronic Science and Technology of China), Shuhang Gu (University of Electronic Science and Technology of China)
CodeRestorationSuper ResolutionTransformerImage
🎯 What it does: This paper proposes a Progressive Focused Transformer (PFT) that enhances feature aggregation in single image super-resolution tasks through Progressive Focused Attention (PFA).
🎯 What it does: A probabilistic OOD classification framework based on class hierarchy, ProHOC, is proposed, which can assign unknown samples to appropriate internal nodes in the hierarchy rather than simply labeling them as OOD.
🎯 What it does: This paper proposes a two-stage movie dubbing generation framework, which first performs acoustic pre-training with voiceprint enhancement, and then freezes the acoustic system to achieve acoustic decoupling of voiceprint and emotional alignment in a non-acoustic mixing manner, thereby generating high-quality and emotionally consistent movie dubbing.
Prototype-Based Image Prompting for Weakly Supervised Histopathological Image Segmentation
Qingchen Tang (University of New South Wales), Yang Song (University of New South Wales)
CodeSegmentationTransformerContrastive LearningImageBiomedical Data
🎯 What it does: A weakly supervised histopathological image segmentation framework based on Prototype-Based Image Prompting (PBIP) is proposed, which constructs a multi-prototype image library using image-level labels and achieves feature matching through contrastive learning, thereby generating more accurate Class Activation Maps (CAM).
🎯 What it does: This paper proposes a ProxUnroll method based on approximate gradients, utilizing proximal trajectory loss to train deep image restorers, enabling them to approach explicit proximal operators in single-pixel imaging (SPI), thus achieving fast and accurate reconstruction.
🎯 What it does: This paper proposes a point cloud representation learning framework PSA-SSL based on self-supervised contrastive learning, which significantly improves the performance of point cloud semantic segmentation and object detection by introducing a 3D bounding box regression pre-training task and LiDAR beam pattern augmentation.
🎯 What it does: A method for detecting backdoor samples in the training set using the prediction shift uncertainty (PSU) generated by Dropout during inference is proposed (PSBD).
🎯 What it does: A pseudo-visible feature fine-grained fusion (PFGF) method is proposed for thermal image object detection, utilizing a prior thermal-to-visible (T2V) translation model to generate pseudo-visible features, which are fused with multi-scale thermal features through a graph neural network.
PVC: Progressive Visual Token Compression for Unified Image and Video Processing in Large Vision-Language Models
Chenyu Yang (Tsinghua University), Jifeng Dai (Tsinghua University)
CodeCompressionTransformerVision Language ModelImageVideoText
🎯 What it does: Proposes Progressive Visual Token Compression (PVC), treating images as 'static videos' to unify the hierarchical compression and encoding of images and videos in the Vision-Language Model.
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageVideoMultimodalityBenchmark
🎯 What it does: A large text-visual content evaluation dataset, Q-Eval-100K, has been proposed, and a unified evaluation model, Q-Eval-Score, has been trained based on this dataset, which can provide separate scores for visual quality and alignment.
🎯 What it does: In the black-box visual prompt learning task, a query-efficient framework based on subspace learning is proposed, utilizing pre-trained models and low-dimensional subspaces to generate input-dependent visual prompts.
R-TPT: Improving Adversarial Robustness of Vision-Language Models through Test-Time Prompt Tuning
Lijun Sheng (University of Science and Technology of China), Ran He (University of Science and Technology of China)
CodeClassificationAdversarial AttackTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: A robust testing prompt tuning method for CLIP, called R-TPT, is proposed to enhance the model's defense capability against adversarial attacks during inference.
🎯 What it does: This paper proposes a retrieval-enhanced geographic location embedding method (RANGE), which utilizes high-resolution satellite image features of similar locations to supplement the high-frequency visual information lost in contrastive learning, thereby generating more expressive multi-scale geographic embeddings.
Rashomon Sets for Prototypical-Part Networks: Editing Interpretable Models in Real-Time
Jon Donnelly (Duke University), Cynthia Rudin (Duke University)
CodeClassificationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: Proposes the Proto-RSet framework, which can quickly generate and interactively use various equivalent performance ProtoPNet models (i.e., the Rashomon set), addressing the interactive bottleneck in model debugging, supporting real-time insertion/deletion of prototypes while ensuring accuracy.
🎯 What it does: This paper proposes RCP-Bench, a comprehensive benchmark for evaluating the robustness of multi-vehicle collaborative perception under various real-world distortions (such as weather, camera failures, and time mismatches), and presents two training strategies, RCP-Drop and RCP-Mix, to enhance robustness for this benchmark.
Realistic Test-Time Adaptation of Vision-Language Models
Maxime Zanella (UCLouvain), Ismail Ben Ayed
CodeDomain AdaptationTransformerVision Language ModelImage
🎯 What it does: This paper proposes a real-time adaptation method for visual language models, Stat A, aimed at maintaining zero-shot performance in more realistic deployment scenarios.
🎯 What it does: A unified scene text editing method RS-STE is proposed, which utilizes the implicit style-content separation of the recognition model to achieve text content replacement while maintaining style.
🎯 What it does: A relationship consistency learning framework named ReCon is proposed to address the noisy correspondence problem in cross-modal retrieval and improve the recognition and retrieval performance of true correspondences.
🎯 What it does: A framework called RDMD is proposed, which unifies the use of a single pre-trained diffusion model for deterministic regression and random sampling in zero-shot image restoration tasks.
Recover and Match: Open-Vocabulary Multi-Label Recognition through Knowledge-Constrained Optimal Transport
Hao Tan (Institute of Automation, Chinese Academy of Sciences), Zhen Lei (Institute of Automation, Chinese Academy of Sciences)
CodeClassificationRecognitionVision Language ModelContrastive LearningImage
🎯 What it does: Proposes the RAM framework, which recovers the local semantics of CLIP through the Ladder Local Adapter and uses Knowledge-Constrained Optimal Transport for precise matching of image regions and labels, achieving open vocabulary multi-label recognition.
🎯 What it does: This paper proposes an algorithm for correcting classifier-free guidance (ReCFG), which addresses the theoretical flaw of expected bias produced by traditional CFG in diffusion model sampling, achieving more accurate conditional sampling without the need for retraining.
Recurrence-Enhanced Vision-and-Language Transformers for Robust Multimodal Document Retrieval
Davide Caffagni (University of Modena and Reggio Emilia), Rita Cucchiara (Istituto Italiano di Tecnologia)
CodeRetrievalRecurrent Neural NetworkTransformerVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: A cross-modal retrieval framework named ReT is proposed, supporting multi-modal queries and documents that include images and text, utilizing multi-layer visual and textual representations for retrieval.
Redefining <Creative> in Dictionary: Towards an Enhanced Semantic Understanding of Creative Generation
Fu Feng (Southeast University), Xin Geng (Southeast University)
CodeGenerationDiffusion modelText
🎯 What it does: Proposes redefining 'creative' as a new token <CreTok>, enabling the diffusion model to directly generate composite creative images in a zero-shot manner through image-free training.
🎯 What it does: This paper proposes an incremental learning method (CREATE) that utilizes lightweight autoencoders to learn class-separating low-dimensional manifolds, aiming to reduce inter-class confusion and alleviate catastrophic forgetting.
Relative Pose Estimation through Affine Corrections of Monocular Depth Priors
Yifan Yu (ETH Zurich), Viktor Larsson (Lund University)
CodePose EstimationDepth EstimationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: A depth prior using a monocular depth estimation model is proposed, and the relative pose of the camera is solved through explicit scale and offset correction.
Reloc3r: Large-Scale Training of Relative Camera Pose Regression for Generalizable, Fast, and Accurate Visual Localization
Siyan Dong (University of Hong Kong), Yanchao Yang (University of Hong Kong)
CodePose EstimationTransformerSimultaneous Localization and MappingImage
🎯 What it does: Proposes the Reloc3r framework, which achieves efficient and generalizable visual localization through a large-scale relative pose regression network and a minimal motion averaging module;
Remote Photoplethysmography in Real-World and Extreme Lighting Scenarios
Hang Shao (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)
CodeTransformerContrastive LearningVideo
🎯 What it does: An end-to-end remote photoplethysmography (rPPG) model based on visual Transformer is proposed, capable of non-contact measurement of physiological indicators such as heart rate under extreme lighting and complex interference scenarios in the real world.
🎯 What it does: A negative embedding learning framework called ReNeg is proposed, which automatically optimizes negative embeddings through gradient descent to enhance the visual quality and human preference of text-to-image/video generation.
ResCLIP: Residual Attention for Training-free Dense Vision-language Inference
Yuhang Yang (University of Electronic Science and Technology of China), Lixin Duan (University of Electronic Science and Technology of China)
CodeSegmentationTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes an untrained framework called ResCLIP, which utilizes the intermediate layer self-attention of the CLIP model to improve dense visual-language inference, particularly for open vocabulary semantic segmentation.
ReSpec: Relevance and Specificity Grounded Online Filtering for Learning on Video-Text Data Streams
Chris Dongjoo Kim (Seoul National University), Christopher Clark (Allen Institute for AI)
CodeRetrievalOptimizationComputational EfficiencyTransformerVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes an online filtering framework named ReSpec, which is designed to real-time select high-quality samples that meet task requirements from video-text data streams to improve online learning efficiency and performance.
Rethinking Decoder Design: Improving Biomarker Segmentation Using Depth-to-Space Restoration and Residual Linear Attention
Saad Wazir (Korea Advanced Institute of Science and Technology), Daeyoung Kim (Korea Advanced Institute of Science and Technology)
CodeSegmentationConvolutional Neural NetworkBiomedical Data
🎯 What it does: A new multi-scale convolutional attention deep-to-space (MCADS) decoder is proposed for biomarker medical image segmentation, achieving high-precision segmentation in conjunction with an improved U2-Net encoder.
Rethinking Personalized Aesthetics Assessment: Employing Physique Aesthetics Assessment as An Exemplification
Haobin Zhong (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)
CodeRecommendation SystemGraph Neural NetworkLarge Language ModelSupervised Fine-TuningGenerative Adversarial NetworkImageMultimodality
🎯 What it does: This paper proposes a new personalized aesthetic assessment framework PAA+, which achieves precise predictions of user personalized aesthetics through three stages (pre-training, fine-tuning, and continuous learning), using physical aesthetics as an experimental case.
Rethinking Query-based Transformer for Continual Image Segmentation
Yuchen Zhu (ShanghaiTech University), Sibei Yang (Sun Yat-sen University)
CodeSegmentationTransformerImage
🎯 What it does: The SimCIS framework is proposed to address the issues of catastrophic forgetting and background semantic drift in class-incremental image segmentation.
🎯 What it does: This paper proposes ProtoGCN, which achieves fine-grained skeleton action recognition through a prototype reconstruction network, allowing for better differentiation of similar actions.
Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift
Siyuan Liang (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
CodeDomain AdaptationAdversarial AttackTransformerLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: The study evaluates the generalization effect of backdoor attacks under domain transfer during the instruction tuning process of visual-language models.
🎯 What it does: A dynamic weighted gradient aggregation method based on performance differences (PIVRG) is proposed for multi-task learning, aiming to reduce performance variance between tasks and enhance overall performance.
🎯 What it does: This paper proposes an image-level generative replay method based on Stable Diffusion and a similarity-based cross-sampling mechanism to prevent catastrophic forgetting in class-incremental object detection.
🎯 What it does: By compressing 3D Gaussian mixture shapes into a learnable small basis, high-fidelity, animatable head avatars are constructed in real-time using FLAME parameters.
🎯 What it does: A camera-radar 3D object detection framework called RICCARDO is proposed, which predicts radar strike distribution and aligns it through convolution to obtain precise locations using monocular detection priors.
🎯 What it does: A vision-based robot pose and joint angle estimation framework called RoboPEPP is proposed, which utilizes self-supervised embedding prediction with joint occlusion to enhance the encoder's understanding of the robot's physical model.
Robust Message Embedding via Attention Flow-Based Steganography
Huayuan Ye (East China Normal University), Chenhui Li (East China Normal University)
CodeData SynthesisCompressionSafty and PrivacyTransformerFlow-based ModelImage
🎯 What it does: A robust information embedding framework called RMSteg based on attention flow is proposed, utilizing reversible QR code transfer, reversible token fusion, and attention coupling networks to achieve high capacity, robustness, and high-quality image steganography.
Robust Multimodal Survival Prediction with Conditional Latent Differentiation Variational AutoEncoder
Junjie Zhou (Nanjing University of Aeronautics and Astronautics), Wei Shao (Nanjing University of Aeronautics and Astronautics)
CodeGenerationData SynthesisRepresentation LearningTransformerAuto EncoderMultimodalityBiomedical Data
🎯 What it does: This study investigates the generation of multifunctional genomic embeddings from pathological images for robust multimodal survival prediction in the context of missing genomic data.
🎯 What it does: A robust object removal model RORem based on semi-supervised learning and human-computer interaction is proposed, and a dataset of approximately 200K object removal pairs is constructed.
🎯 What it does: To achieve high-quality interactive segmentation of moving objects in remote sensing videos, the ROS-SAM method is proposed, along with a specially designed data and inference pipeline.
Rotation-Equivariant Self-Supervised Method in Image Denoising
Hanze Liu (Xi'an Jiaotong University), Deyu Meng (Macau University of Science and Technology)
CodeRestorationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a self-supervised image denoising network called AdaReNet, which integrates rotational equivariance priors into the self-supervised denoising framework for the first time, and theoretically analyzes the impact of upsampling and downsampling on equivariance within the U-Net structure.
RSAR: Restricted State Angle Resolver and Rotated SAR Benchmark
Xin Zhang (Nankai University), Xiang Li (Nankai University)
CodeObject DetectionImageBenchmark
🎯 What it does: To address the issue of discontinuity in angle prediction boundaries in rotating object detection, a Unit Circle Constrained Angle Resolver (UCR) is proposed, which enhances the angle prediction accuracy of weakly supervised models. Based on this, the first large-scale multi-class rotating SAR object detection dataset, RSAR, is constructed.
🎯 What it does: Proposed the S2D-LFE method, which utilizes sparse perspective light field events (LFE) to generate dense, spatiotemporally consistent new perspective LFEs;
🎯 What it does: This paper proposes a 3D medical image registration network named SACB-Net, which utilizes Spatially Aware Convolution Blocks (SACB) to adaptively generate convolution kernels, thereby enhancing feature representation in different spatial regions.
🎯 What it does: By upgrading the pre-trained Segment Anything Model (SAM) from image to video, SAM-I2V is proposed to support Promptable video segmentation.
🎯 What it does: A unified framework called Samba based on Mamba is proposed to handle various salient object detection tasks, including RGB/RGB-D/RGB-T, video SOD, and RGB-D video SOD;
Sample- and Parameter-Efficient Auto-Regressive Image Models
Elad Amrani (Apple), Alex Bronstein (Technion)
CodeRecognitionGenerationTransformerImage
🎯 What it does: We propose XTRA, a self-regressive image model based on Vision Transformer, which predicts pixels block by block using a Block Causal Mask;
🎯 What it does: This paper proposes a Sampling Innovation-based Adaptive Compressed Sensing (SIB-ACS) framework, which combines multi-stage negative feedback adaptive sampling with a novel reconstruction network PCCD-Net to achieve high-fidelity image reconstruction.
🎯 What it does: This paper proposes an open set recognition method for point clouds called SASep, which decomposes objects using semantic saliency and distinguishes between known and unknown categories through geometric synthesis and feature distance enhancement.
SATA: Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers
Nick Nikzad (Griffith University), Jun Zhou (Griffith University)
CodeClassificationTransformerImage
🎯 What it does: Without the need for additional training, Spatial Autocorrelation Token Analysis (SATA) is introduced, enhancing the model's representation ability and robustness by segmenting and merging the spatial autocorrelation of ViT feature tokens.
Satellite Observations Guided Diffusion Model for Accurate Meteorological States at Arbitrary Resolution
Siwei Tu (Fudan University), Lei Bai (Fudan University)
CodeGenerationData SynthesisOptimizationDiffusion modelTime Series
🎯 What it does: A diffusion model conditioned on satellite observations (SGD) has been constructed, which generates high-resolution meteorological fields at a scale of 6.25 km by using low-resolution ERA5 images, GridSat satellite data, and meteorological station observations, achieving precise downscaling of meteorological states.
Scalable Video-to-Dataset Generation for Cross-Platform Mobile Agents
Yunseok Jang (University of Michigan), Honglak Lee (University of Michigan)
CodeObject DetectionData SynthesisRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningVideoMultimodality
🎯 What it does: An automated process was developed to extract and annotate mobile operation workflows from YouTube tutorial videos, resulting in the creation of the cross-platform (iOS/Android) mobile OS navigation dataset MONDAY, which includes 20K videos and 313K frames. Pre-training and fine-tuning of models were conducted on this dataset.
🎯 What it does: A dynamic sample pruning framework called SeTa is proposed, which achieves training acceleration on large-scale datasets through random downsampling, loss-based clustering, and a sliding window strategy, while reducing training costs without compromising or even enhancing model performance.
🎯 What it does: A block and patch compression tokenization (BPT) method is proposed, which reduces the length of mesh sequences by approximately 75%, thereby supporting the generation of over 8k face meshes.
CodeDomain AdaptationPrompt EngineeringVision Language ModelImageText
🎯 What it does: To address the issue of adaptive testing of visual-language models under domain shift conditions, a Supportive Clique Attribute Prompt (SCAP) framework is proposed, enabling cross-sample information fusion and adaptation during transductive testing.
Scene Map-based Prompt Tuning for Navigation Instruction Generation
Sheng Fan (Zhejiang University), Yi Yang (Zhejiang University)
CodeGenerationGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringPoint Cloud
🎯 What it does: A navigation instruction generation framework called MAPINSTRUCTOR is proposed, which is based on scene map prompt tuning. It combines 3D voxel scene encoding, global topological map prompts, and landmark uncertainty assessment to achieve more accurate navigation instruction generation.
SceneTAP: Scene-Coherent Typographic Adversarial Planner against Vision-Language Models in Real-World Environments
Yue Cao (Agency for Science Technology and Research), Qing Guo (Agency for Science Technology and Research)
CodeAdversarial AttackTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: A new scene-consistent typography adversarial attack method (SceneTAP) is proposed, aimed at misleading large visual-language models (LVLMs) while maintaining visual naturalness.
🎯 What it does: This paper proposes a Time Prediction Diffusion Model (TPDM), which incorporates a Time Prediction Module (TPM) into the diffusion model to dynamically predict the noise level and total sampling steps for each instance, thereby improving image quality and sampling efficiency.
🎯 What it does: A multi-modal 3D semantic occupancy prediction framework SDG-OCC is proposed, which combines camera and LiDAR information to achieve more accurate occupancy predictions.
🎯 What it does: This paper proposes a Self-Calibrating Adaptive Prompt (SeCap) framework to address the challenge of cross-view person identification between drones and ground cameras.
Seeing the Abstract: Translating the Abstract Language for Vision Language Models
Davide Talon (Fondazione Bruno Kessler), Yiming Wang
CodeRetrievalTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: This study addresses the issue of the lack of expression of abstract language in visual language models within the fashion domain and proposes a training-free, model-agnostic abstract-to-concrete translation method (ACT) that significantly improves retrieval performance through language rewriting and representation offset compensation.
🎯 What it does: This paper proposes CLIP-PGS, a progressive generation-selection masking strategy aimed at enhancing the training efficiency of CLIP in visual-language pre-training while preserving key information.
🎯 What it does: A semi-supervised image-text sentiment recognition framework named SCRD is proposed, which significantly improves sentiment recognition performance in label-scarce environments by utilizing techniques such as feature decoupling (separating common and private features), unimodal classifiers, modality selection attention (MSeA), and pseudo-label filtering (PLF).
🎯 What it does: A motion-guided reconstruction method for color spike cameras (CSC) is proposed, and a self-supervised network is built based on this to remove quantization noise, ultimately achieving high-quality color image recovery from Bayer-pattern spike flows.
🎯 What it does: This paper presents SemanticDraw, a system that supports real-time interaction and allows for the generation of high-quality images through hand-drawn semantic masks controlling multiple text prompts on a canvas.
SfM-Free 3D Gaussian Splatting via Hierarchical Training
Bo Ji (National University of Singapore), Angela Yao (National University of Singapore)
CodeGenerationPose EstimationGaussian SplattingSimultaneous Localization and MappingVideoPoint Cloud
🎯 What it does: A 3D Gaussian spraying method without SfM preprocessing is proposed, which combines basic 3DGS models of different scene segments into a complete scene model using hierarchical training.
🎯 What it does: Generate Spherical Gaussians from 2D image edge information and use them to extract 3D parametric curves, achieving efficient 3D curve reconstruction.
🎯 What it does: This paper proposes SGFormer, a 3D semantic scene completion framework that integrates satellite images with ground camera views, addressing the occlusion bottleneck that arises from relying solely on ground perspectives.
🎯 What it does: The research uses geometric priors to guide the diffusion model in generating synthetic image shadows, significantly improving the geometric quality of the shadows.
🎯 What it does: This paper constructs the Flow-R dataset by modifying the shape and texture of target objects and adding unseen objects to the original KITTI images, aimed at evaluating the robustness of optical flow estimation.
🎯 What it does: An end-to-end visual-language-action model called ShowUI is proposed and trained for executing localization and navigation tasks in GUI environments.
Silence is Golden: Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation
Yuan Gan (Zhejiang University), Yi Yang (Zhejiang University)
CodeGenerationSafty and PrivacyAdversarial AttackDiffusion modelVideoAudio
🎯 What it does: A two-stage active protection method called Silencer is proposed to generate robust adversarial perturbations in audio-driven speaker animation models based on LDM, thereby protecting portrait privacy.
🎯 What it does: This paper studies causal representation in multi-agent interactions, proposing a regularization method to enhance the model's perception of causal relationships and achieve causal transfer from simulation to real scenarios.
Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking
Chaocan Xue (Guangxi Normal University), Shuxiang Song (Guangxi Normal University)
CodeObject TrackingTransformerVideo
🎯 What it does: This study investigates the layer redundancy issue in lightweight ViT trackers and proposes a similarity-guided layer adaptive mechanism that selects a single representative layer to prune redundant layers, achieving real-time drone target tracking.
CodeDomain AdaptationKnowledge DistillationPrompt EngineeringVision Language ModelImage
🎯 What it does: A few-shot counting model for single-source domain generalization, URM, is proposed, which achieves cross-domain counting through distilled universal visual-language prototypes from CLIP.
🎯 What it does: A Sketchy-3DIS framework is proposed to achieve weakly supervised 3D instance segmentation with only imprecise sketchy bounding boxes provided.
Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves
Shihan Wu (University of Electronic Science and Technology of China), Heng Tao Shen (Tongji University)
CodeClassificationDomain AdaptationComputational EfficiencyTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: This paper proposes the Skip Tuning method, which performs layer skipping (LSkip) and category skipping (CSkip) directly on the CLIP pre-trained model to achieve transfer learning for downstream tasks without adding extra prompt vectors or adapters.
SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos
Yuzheng Liu (Peking University), Baoquan Chen (Peking University)
CodePose EstimationDepth EstimationRobotic IntelligenceTransformerSimultaneous Localization and MappingVideoPoint Cloud
🎯 What it does: SLAM3R is proposed, a system that achieves real-time, high-definition dense 3D reconstruction using RGB video, employing a two-layer neural network framework for end-to-end point cloud prediction and global registration;
🎯 What it does: Proposes the SMILE framework, which uses high-level semantic features from the CLIP pre-trained model and synthetic motion data for self-supervised learning in a video mask autoencoder.