European Conference on Computer Vision · 980 papers
Reason2Drive: Towards Interpretable and Chain-based Reasoning for Autonomous Driving
Ming Nie (Fudan University), Li Zhang (Fudan University)
CodeAutonomous DrivingExplainability and InterpretabilityTransformerVision Language ModelVideoTextChain-of-Thought
🎯 What it does: This paper proposes an autonomous driving framework based on interpretable chain reasoning, constructs a large-scale dataset named Reason2Drive, and designs a specialized evaluation metric called ADRScore for chain reasoning, further improving visual language models (VLMs) to better utilize visual prior information.
Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch
Taemin Park (KAIST), Heeyoung Kim (KAIST)
CodeClassificationImage
🎯 What it does: Propose a semi-supervised learning framework RECD that addresses label imbalance and distribution mismatch between unlabeled and labeled data by estimating the class distribution of unlabeled data to rebalance the model.
🎯 What it does: The paper proposes a lightweight concept elimination method called Receler, which reliably removes specified concepts from text-to-image diffusion models without accessing image data.
🎯 What it does: Propose a novel recursive temporal fusion framework called RecurrentBEV for bird's-eye-view 3D object detection using multi-view cameras.
Jiaxin Ge (UC Berkeley), Trevor Darrell (UC Berkeley)
CodeAI Code AssistantTransformerLarge Language ModelPrompt EngineeringImageVideoMultimodality
🎯 What it does: Propose a Recursive Visual Programming (RVP) approach, enabling Large Language Models (LLMs) to generate modular, dynamically typed code through recursive self-calls in Visual Question Answering (VQA) tasks, progressively decomposing complex problems layer by layer;
🎯 What it does: Designed and implemented an end-to-end event camera reconstruction model named REDIR, which can accomplish event alignment, filtering, and reconstruction in occlusion-free event-based synthetic aperture imaging (E-SAI), ultimately generating high-resolution images without occlusion.
🎯 What it does: This paper proposes an end-to-end learning-based image compression framework named Segmentation-Prior-Guided Image Compression (SegPIC), which utilizes class-agnostic semantic masks during training to guide the network to generate region-adaptive transformations, thereby improving pixel-level compression quality.
🎯 What it does: By converting the multi-task sparse supervision learning problem into regional-level distribution comparison, this paper achieves cross-task consistency among different tasks.
🎯 What it does: Propose a region-based visual tokenization method called Reader, constructing an autoencoder to achieve efficient encoding and local editing.
Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density
Peiyu Yang (University of Western Australia), Ajmal Mian (University of Western Australia)
CodeClassificationExplainability and InterpretabilityAdversarial AttackImage
🎯 What it does: This paper proposes a regularization method that utilizes the gradient of the marginal density of input samples to smooth out, reducing the model's reliance on non-robust features, and distinguishes robust from non-robust features through gradient consistency.
🎯 What it does: Proposes a cross-modal contrastive learning framework called CLIP4MC, which trains an RL-friendly vision-language model as a task reward function and constructs a high-quality Minecraft YouTube dataset.
Nico Messikommer (University of Zurich), Davide Scaramuzza (University of Zurich)
CodeTransformerReinforcement LearningSimultaneous Localization and MappingImageVideo
🎯 What it does: Model visual odometry (VO) as a sequence decision problem, training an agent network using reinforcement learning to dynamically adjust parameters such as keyframe selection and grid size, thereby improving the accuracy and robustness of VO.
ReMamber: Referring Image Segmentation with Mamba Twister
Yuhuan Yang (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)
CodeSegmentationVision Language ModelMultimodality
🎯 What it does: Proposed the ReMamber model, which achieves efficient image-text fusion in Referring Image Segmentation (RIS) tasks by utilizing the Mamba Twister block.
🎯 What it does: Propose ReMatching: first perform geometry-preserving low-resolution reconstruction, then conduct functional mapping on the low-resolution model, and finally project the results back to the original high resolution, achieving scalable correspondence for shapes with millions of vertices.
🎯 What it does: Propose Token Adapter, which compresses Vision Transformer tokens by deleting entire rows and columns of the image feature map, achieving acceleration for dense prediction without training.
RePOSE: 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency
Ziming Sun (South China University of Technology), Shengfeng He (Singapore Management University)
CodePose EstimationTransformerVideo
🎯 What it does: Proposes a method called RePOSE, which addresses occlusion problems in video-based 3D human pose estimation by introducing spatial-temporal relative depth consistency supervision.
🎯 What it does: Proposes a domain generalization framework RES based on the bias-variance decomposition perspective, which includes a stabilization module with feature frequency domain enhancement and parameter mutual fusion;
Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures
Jiaxing Huang (Institute of Automation, Chinese Academy of Sciences), Ge Yang (Institute of Automation, Chinese Academy of Sciences)
CodeSegmentationConvolutional Neural NetworkBiomedical Data
🎯 What it does: Accurate segmentation of long, thin tubular structures was achieved by combining pixel-level fractal feature maps (FFM) with a multi-decoder network (MD-Net).
🎯 What it does: This paper proposes a unified vector field representation, RepVF, to integrate multiple tasks such as 3D object detection and 3D lane detection into a single framework, and constructs the RFTR network to achieve single-head multi-task learning, significantly reducing computational redundancy and task competition;
🎯 What it does: Designed a source-free continuous test-time adaptation framework based on uncertainty-adaptive buffering and graph structure preservation, which can efficiently collect reliable samples and prevent catastrophic forgetting and error accumulation under online unsupervised environments.
🎯 What it does: This paper systematically evaluates the robustness of entropy coding models in distributed deep networks and proposes a defense method based on decoupling in frequency and spatial domains.
CodePose EstimationDepth EstimationSimultaneous Localization and MappingImage
🎯 What it does: Automatically resolve scale ambiguity in structure from motion and multi-view 3D reconstruction by analyzing the size of focal blur from multi-view images captured using dual-pixel (Dual-Pixel, DP) sensors.
Minheng Ni (Hong Kong Polytechnic University), Wangmeng Zuo (Harbin Institute of Technology)
CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelDiffusion modelImageChain-of-Thought
🎯 What it does: Proposes the Responsible Visual Editing task, which leverages a multimodal model to automatically identify and edit risky concepts in images, reducing the need for human manual intervention.
🎯 What it does: Proposed RAM, a universal blind image restoration framework based on masked image modeling, which includes a two-phase strategy of mask pre-training and fine-tuning only key layers.
Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework
Wei Suo (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
CodeSegmentationReinforcement LearningPrompt EngineeringVision Language ModelImage
🎯 What it does: Investigated the impact of visual prompt selection in In-Context Learning (ICL) segmentation, and proposed a Stepwise Context Search (SCS) method based on clustering and reinforcement learning to achieve automated high-quality example selection;
Rethinking Data Bias: Dataset Copyright Protection via Embedding Class-wise Hidden Bias
Jinhyeok Jang (Electronics And Telecommunications Research Institute), Chan-Hyun Youn (Electronics And Telecommunications Research Institute)
CodeSafty and PrivacyAuto EncoderImage
🎯 What it does: Proposed a dataset watermarking technique called 'undercover bias,' which embeds invisible hidden biases (i.e., watermarks) into each category of the target dataset, enabling trained models to recognize and classify these watermarks, thereby verifying unauthorized data usage in black-box scenarios.
🎯 What it does: To accelerate MRI reconstruction, this paper proposes an adaptive gradient and momentum accelerated deep unfolding model, achieving more efficient and accurate multi-coil reconstruction through multi-slice related low-memory sensitivity map estimation.
Rethinking Fast Adversarial Training: A Splitting Technique To Overcome Catastrophic Overfitting
Masoumeh Zareapoor (East China Normal University), Pourya Shamsolmoali (Queen's University Belfast)
CodeClassificationAdversarial AttackImage
🎯 What it does: The paper proposes a fast adversarial training framework (RAT) based on the Douglas-Rachford splitting technique, aiming to effectively avoid catastrophic overfitting by stabilizing training dynamics.
🎯 What it does: Designed and verified an Independent Hierarchical Pyramid (IHP) without feature fusion, proposed Soft Nearest Neighbor Interpolation (SNI) and Extended Spatial Window Adaptive Downsampling (ESD), improved the lightweight GSConv (GSConvE), and integrated these techniques into a secondary feature alignment (SA) scheme for real-time detection.
🎯 What it does: This paper proposes a new evaluation metric gAcc and a lightweight feature refinement (FR) module based on Vision Transformer to enhance the performance of few-shot class-incremental learning (FSCIL).
Rethinking Image Super Resolution from Training Data Perspectives
Go Ohtani (Keio University), Yoshimitsu Aoki (Keio University)
CodeSuper ResolutionImage
🎯 What it does: Proposes an automated image evaluation pipeline, constructing the DiverSeg dataset, which is low-resolution but high-quality and object-diverse, for training image super-resolution models.
🎯 What it does: Developed a Density Discriminative Feature Embedding (DDFE) module, achieving cross-domain semantic segmentation model generalization by mining multi-density distributions in single-source LiDAR point clouds, addressing density discrepancy issues caused by different LiDAR sensors.
🎯 What it does: Propose a novel DG-ReID framework named ReNorm, which employs two forward passes respectively adopting Remix Normalization (RN) and Emulation Normalization (EN), while introducing a Domain Frozen (DF) mechanism in both to suppress overfitting of normalization layers on the source domain.
Rethinking Unsupervised Outlier Detection via Multiple Thresholding
Zhonghang Liu (Singapore Management University), Wen-Yan Lin (Singapore Management University)
CodeAnomaly DetectionImage
🎯 What it does: Propose a multi-threshold (Multi-T) module that automatically generates two thresholds using unlabeled data, classifying samples into clean inliers and outliers, thereby improving the scores and annotation effectiveness of unsupervised anomaly detection.
REVISION: Rendering Tools Enable Spatial Fidelity in Vision-Language Models
Agneet Chatterjee (Arizona State University), Chitta R Baral
CodeImage TranslationGenerationData SynthesisVision Language ModelDiffusion modelScore-based ModelImageTextMultimodalityMeshBenchmark
🎯 What it does: Built REVISION, a 3D rendering pipeline based on Blender, capable of precisely synthesizing high-quality images containing over 100 3D assets, 11 types of spatial relationships, diverse backgrounds, and perspectives. These synthetic images are utilized to enhance the spatial consistency of text-image models without training, while introducing the RevQA evaluation framework to assess the spatial reasoning capabilities of multimodal large language models.
Revisit Anything: Visual Place Recognition via Image Segment Retrieval
Kartik Garg (Indian Institute of Science), Sourav Garg (University of Adelaide)
CodeSegmentationRetrievalImage
🎯 What it does: By leveraging open-set segmentation for visual place recognition to obtain SuperSegment, and performing feature aggregation and retrieval at the segment level, the robustness to viewpoint changes is enhanced;
Revisiting Adaptive Cellular Recognition Under Domain Shifts: A Contextual Correspondence View
Jianan Fan (University of Sydney), Weidong Cai (Microsoft)
CodeSegmentationDomain AdaptationKnowledge DistillationConvolutional Neural NetworkTransformerGenerative Adversarial NetworkContrastive LearningImageBiomedical Data
🎯 What it does: This paper proposes a cross-domain cell nucleus recognition framework based on biological context correspondence, which learns high-level pathological generation principles from implicit correspondence relationships between cells and tissues, as well as between cells, through unsupervised self-supervised tasks, thereby achieving domain adaptation.
🎯 What it does: A generation-composition framework is constructed for cross-domain adverse weather object detection, with core components including IAoU loss improvement for regression, joint filtering combined with student perception for pseudo-label screening, and image enhancement based on rendering/recovery and super-resolution.
Seongbo Ha (Sungkyunkwan University), Hyeonwoo Yu (Sungkyunkwan University)
CodeGaussian SplattingSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Propose a real-time dense representation SLAM framework that integrates G-ICP with 3D Gaussian Splatting (3DGS), which simultaneously performs tracking and mapping on a single Gaussian map, achieving a high frame rate of 107 FPS and excellent map quality.
CodeRetrievalTransformerVision Language ModelContrastive LearningVideo
🎯 What it does: Proposes an RGNet, a unified long video text retrieval and localization network, achieving end-to-end retrieval and localization of specified events in 20-120 minute long videos.
🎯 What it does: Propose a rotation-invariant surface attention-enhanced convolution (RISurConv) based on local triangular face construction for 3D point cloud classification and segmentation tasks.
🎯 What it does: This paper investigates adaptation methods for CLIP in out-of-distribution (OOD) scenarios (Prompt Learning, Adapters, Test-Time Prompt Tuning) and finds that while they improve accuracy, they severely lose calibration. To address this issue, the authors propose three calibration strategies based on logit range normalization (ZS-Norm, Penalty, SaLS).
🎯 What it does: Propose DMRNet, which models representations of different modality combinations as probability distributions and decouples training and inference representations, thereby alleviating the intra-class directional constraints of traditional subspace methods to achieve robust multi-modal learning;
🎯 What it does: Enhancing model robustness against adversarial attacks by learning a small number of private additional tokens (Robustness Tokens) on the Transformer architecture, while maintaining downstream task performance.
🎯 What it does: This paper studies and verifies the application of Rotating Position Embedding (RoPE) in Vision Transformer (ViT), and proposes axial and mixed learning frequency implementations for 2D RoPE.
🎯 What it does: Proposes a robust point-based neural rendering framework, RPBG, for achieving high-quality view synthesis in diverse outdoor scenarios.
🎯 What it does: Studied a method to directly learn NeRF from rolling shutter images, jointly optimizing camera trajectory and voxel networks to achieve rolling shutter distortion compensation and novel view synthesis.
🎯 What it does: This paper proposes the first rolling shutter bundle adjustment (RSL-BA) framework based on line features, deriving the curve projection under rolling shutter using Pucker line parameterization and constructing a stable projection error;
SA-DVAE: Improving Zero-Shot Skeleton-Based Action Recognition by Disentangled Variational Autoencoders
Sheng-Wei Li (National Taiwan University), Jane Yung-jen Hsu (National Taiwan University)
CodeRecognitionAuto EncoderGraph
🎯 What it does: This paper proposes the SA-DVAE model, which improves zero-shot recognition of skeletal actions through feature separation and alignment.
Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models
Samuele Poppi (University of Modena and Reggio Emilia), Rita Cucchiara (University of Modena and Reggio Emilia)
CodeGenerationData SynthesisRetrievalSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelDiffusion modelContrastive LearningMultimodality
🎯 What it does: Designed and implemented Safe-CLIP, which fine-tunes the text and visual encoders of CLIP to be insensitive to NSFW content in cross-modal retrieval, text-to-image, and image-to-text generation tasks.
Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion
Sanghyun Kim (Korea Advanced Institute of Science and Technology), Juho Lee (Korea Advanced Institute of Science and Technology)
CodeGenerationSafty and PrivacyReinforcement LearningPrompt EngineeringDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the Human Feedback Inversion (HFI) framework, which compresses human feedback into soft word embeddings to eliminate harmful or copyright concepts in text-to-image diffusion models, and combines it with Safe Self-Distillation Diffusion (SDD) to further fine-tune the model and enhance concept removal effectiveness.
🎯 What it does: Propose a multi-exposure HDR image reconstruction network named SAFNet, which achieves fast ghost removal and high-quality HDR synthesis through selective alignment and explicit fusion.
SAH-SCI: Self-Supervised Adapter for Efficient Hyperspectral Snapshot Compressive Imaging
Haijin Zeng (IMEC-Ghent University), Jingyong Su (Harbin Institute of Technology (Shenzhen))
CodeCompressionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a self-supervised adapter (SAH), achieving unsupervised fine-tuning for hyperspectral snapshot compression imaging by adding a lightweight adaptation module after a frozen pre-trained model.
🎯 What it does: Proposes the Salience-Based Adaptive Masking (SBAM) method, which intelligently selects mask positions by calculating the salience of image tokens and achieves adaptive pre-training for each image through dynamic mask ratio (AMR).
Scalar Function Topology Divergence: Comparing Topology of 3D Objects
Ilya Trofimov (Skolkovo Institute of Science and Technology), Serguei Barannikov (Skolkovo Institute of Science and Technology)
CodeSegmentationMeshGraphBiomedical Data
🎯 What it does: Proposed and implemented a new topology comparison tool called Scalar Function Topology Divergence (SFTD), along with its corresponding F-Cross-Barcode, to measure and locate multi-scale topological differences between sublevel sets of two scalar functions, and applied SFTD as a loss function for 3D shape reconstruction and segmentation tasks.
🎯 What it does: Propose Asynchronous Score Distillation (ASD), achieving scalable text-to-3D synthesis by reducing noise prediction error through time-step advancement without fine-tuning diffusion model weights.
Ryo Nakamura (National Institute of Advanced Industrial Science and Technology), Hirokatsu Kataoka (National Institute of Advanced Industrial Science and Technology)
🎯 What it does: Propose using a single fractal image with noise transformation to construct a minimal synthetic pre-training dataset called 1p-frac in visual pre-training, and achieve model pre-training through LPCE loss.
🎯 What it does: This paper proposes a deep learning framework called ScanTalk, which can generate animations of arbitrary topology 3D facial meshes driven by speech, without requiring pre-registration of the mesh.
🎯 What it does: Propose ScatterFormer, a voxel transformer utilizing hash-based linear attention (SLA) and cross-window interaction (CWI) for large-scale point cloud 3D detection, significantly reducing computational and memory overhead under sparse point clouds.
CodeGenerationData SynthesisRetrievalTransformerLarge Language ModelGaussian SplattingTextMeshRetrieval-Augmented Generation
🎯 What it does: Propose SceneTeller, an end-to-end pipeline that generates complete 3D room scenes from natural language descriptions and supports style editing for scenes or individual objects.
CodeOptimizationSafty and PrivacyComputational EfficiencyData-Centric LearningImageMultimodality
🎯 What it does: Propose a novel machine unlearning method called Scissorhands, which can effectively remove the impact of specified data while maintaining the model's performance on the remaining data.
SCLIP: Rethinking Self-Attention for Dense Vision-Language Inference
Feng Wang (Johns Hopkins University), Alan Yuille (Johns Hopkins University)
CodeSegmentationTransformerPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: By improving CLIP's self-attention mechanism, the SCLIP model is proposed, achieving open-source semantic segmentation without additional training.
SCOMatch: Alleviating Overtrusting in Open-set Semi-supervised Learning
Zerun Wang (University of Tokyo), Toshihiko Yamasaki (University of Tokyo)
CodeClassificationImageBenchmark
🎯 What it does: To address the problem of overfitting on the decision boundary caused by excessive trust in labeled ID samples in open-ended semi-supervised learning, the SCOMatch method is proposed, treating OOD as an additional class to construct a (K+1)-class SSL.
SDPT: Synchronous Dual Prompt Tuning for Fusion-based Visual-Language Pre-trained Models
Yang Zhou (Beihang University), Yan Xu (Zhejiang University)
CodeObject DetectionTransformerPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: This paper proposes a synchronous dual-modal prompt tuning method called SDPT for fused vision-language pre-training models (e.g., GLIP), aiming to achieve parameter-efficient fine-tuning for downstream tasks.
🎯 What it does: Propose a 3D object detection framework SEED based on DETR, achieving high-quality queries and efficient feature interaction on point clouds through Dual Query Selection (DQS) and Deformable Grid Attention (DGA).
Seeing the Unseen: A Frequency Prompt Guided Transformer for Image Restoration
Shihao Zhou (Nankai University), Jufeng Yang (Nankai University)
CodeRestorationTransformerPrompt EngineeringImage
🎯 What it does: Propose a frequency-prompt-based Transformer model called FPro, which decomposes features into low-frequency and high-frequency components through gated dynamic decoupling, and uses dual-frequency prompt modules to separately generate and modulate prompts for image restoration;
🎯 What it does: Propose a self-supervised scene flow estimation method called SeFlow, which improves the accuracy of LiDAR point cloud scene flow by utilizing dynamic point classification and clustering consistency constraints.
🎯 What it does: Propose SegIC, an end-to-end context segmentation framework based on vision foundation models, achieving segmentation from few examples to target images by leveraging dense correspondence relationships.
SegVG: Transferring Object Bounding Box to Segmentation for Visual Grounding
Weitai Kang (Illinois Institute of Technology), Yan Yan (Cisco Research)
CodeRecognitionSegmentationTransformerVision Language ModelMultimodality
🎯 What it does: Propose the SegVG method, which converts bounding box annotations into pixel-level segmentation supervision to improve visual localization tasks.
🎯 What it does: Building upon SeI_T, this paper proposes a self-supervised pre-training scheme called Masked Token Modeling (MTM), combined with two novel token-level augmentation methods, TokenAdapt and ColorAdapt, to construct a complete storage-efficient visual training framework (SeiT++).
🎯 What it does: Propose the SelEx method, combining unsupervised and supervised self-expertise techniques, generating multi-level pseudo labels through hierarchical semi-supervised k-means to achieve a fine-grained general category discovery (GCD) framework for simultaneously discovering and classifying known and unknown categories.
Self-Adapting Large Visual-Language Models to Edge Devices across Visual Modalities
Kaiwen Cai (University College London), Chris Xiaoxuan Lu (Cisco Research)
CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Propose the EdgeVL framework, which can migrate large-scale vision-language (VL) models (e.g., CLIP) to edge devices without manual annotation, and is compatible with RGB and non-RGB (e.g., depth, infrared) multimodal inputs, supporting open-vocabulary classification.
Self-Guided Generation of Minority Samples Using Diffusion Models
Soobin Um (KAIST), Jong Chul Ye (KAIST)
CodeData SynthesisDiffusion modelImage
🎯 What it does: Propose a self-supervised minority class sample generation method that uses only pre-trained diffusion models, leveraging self-guided sampling to induce generation of samples in low-density regions.
🎯 What it does: Propose a two-stage self-supervised co-salient object detection method named SCoSPARC, leveraging the correspondence between local and global scale features of ViT to segment co-occurring salient objects in image groups.
🎯 What it does: Proposes a self-supervised local-to-global feature adaptation framework (LSFA) to adapt features from pre-trained models to 3D industrial defect detection tasks, improving detection performance through cross-modal alignment and single-modal compression.
Self-supervised Shape Completion via Involution and Implicit Correspondences
Mengya Liu (ETH Zurich), Federico Tombari (Google)
CodeRestorationGenerationPoint Cloud
🎯 What it does: Propose a self-supervised 3D shape completion framework that leverages shape correspondence and self-inverse function constraints to recover complete geometry from incomplete scans.
🎯 What it does: Proposed a self-supervised video duplication localization framework that extracts local features by introducing Regional Tokens in Vision Transformers and trains a temporal localization model using self-supervised generated video pairs
Self-Supervised Video Desmoking for Laparoscopic Surgery
Renlong Wu (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
CodeRestorationConvolutional Neural NetworkTransformerOptical FlowVideoBiomedical Data
🎯 What it does: Proposes a self-supervised surgical video defogging (SelfSVD) method that utilizes previous fog-free frames in the video as unaligned supervision and reference, enabling training and online inference without paired data.
🎯 What it does: This paper proposes an action-aware self-supervised learning method (AA-SSL) that leverages action information generated through object interaction. By aligning action embeddings with corresponding image embeddings, the method enhances the category generalization capability of visual representations.
🎯 What it does: Propose the SelfGeo method, which utilizes self-supervised learning to estimate repeatable, semantically consistent 3D keypoints in point cloud sequences, ensuring the relative positions of keypoints remain unchanged during non-rigid deformations through geodesic distance constraints.
Semantic Diversity-aware Prototype-based Learning for Unbiased Scene Graph Generation
Jaehyeong Jeon (Korea Advanced Institute of Science and Technology), Chanyoung Park (Korea Advanced Institute of Science and Technology)
CodeGenerationMultimodality
🎯 What it does: This paper proposes a model-agnostic semantic diversity-aware prototype learning framework (DPL) for scene graph generation tasks, which captures multiple semantics of the same predicate by learning prototypes and their distributions in the semantic space to achieve unbiased predictions.
Martin Menabue (University of Modena and Reggio Emilia), SIMONE CALDERARA (University of Modena and Reggio Emilia)
CodeClassificationTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: Propose a two-layer semantic residual prompt continual learning method called STAR-Prompt, which utilizes a frozen CLIP text encoder to generate class prototypes and employs them as keys to retrieve the second-layer prompt, injecting semantic residuals into the frozen ViT, balancing model stability and plasticity.
Semantically Guided Representation Learning For Action Anticipation
Anxhelo Diko (Sapienza University of Rome), Luigi Cinque (Sapienza University of Rome)
CodeRecognitionRepresentation LearningTransformerLarge Language ModelVision-Language-Action ModelContrastive LearningVideoMultimodality
🎯 What it does: This paper proposes a framework named S-GEAR for action prediction by learning visual action prototypes and combining them with a language model.
🎯 What it does: This paper proposes a semi-supervised segmentation framework (TopoSemiSeg), which enables the model to learn correct topological structures on unlabeled images through a noise-aware topological consistency loss, thereby improving accuracy in dense distribution gland/cell nucleus segmentation tasks.
🎯 What it does: Proposes a semi-supervised teacher-reference-student framework to learn in action quality assessment (AQA) tasks using only a small amount of labeled data and leveraging a large amount of unlabeled data.
Semi-Supervised Video Desnowing Network via Temporal Decoupling Experts and Distribution-Driven Contrastive Regularization
Hongtao Wu (Hong Kong University of Science and Technology), Lei Zhu (Hong Kong University of Science and Technology)
CodeRestorationConvolutional Neural NetworkTransformerMixture of ExpertsContrastive LearningVideo
🎯 What it does: Propose a semi-supervised video de-snowing network, SemiVDN, which leverages unlabeled real snow videos to enhance generalization and introduces temporal expert modules and distribution-driven contrastive regularization;
Semicalibrated Relative Pose from an Affine Correspondence and Monodepth
Petr Hruby (ETH Zürich), Daniel Barath (ETH Zürich)
CodePose EstimationDepth EstimationImage
🎯 What it does: This paper proposes a method for semi-calibrated relative pose estimation using a single affine correspondence and monocular depth prediction.
🎯 What it does: Proposes the Sparse Focal Point Network (SFPNet), which replaces window attention with a sparse focal point modulation (SFPM) module for semantic segmentation of various types of LiDAR point clouds.
🎯 What it does: Propose a joint optimization framework called SG-NeRF based on scene graphs, which trains neural radiance fields (NeRF) while simultaneously optimizing camera poses and confidence levels, achieving high-quality 3D surface reconstruction even in the presence of significant camera pose noise.
🎯 What it does: This paper proposes the Shape2Scene method, which pretrains 3D scene representations using shape data and performs unsupervised learning on multi-scale high-resolution networks (MH-P, MH-V).
🎯 What it does: Proposes a novel resource-constrained transfer learning framework called SHERL, which employs a two-phase separation strategy of early aggregation and late regulation, significantly reducing GPU memory demand during fine-tuning while maintaining or improving performance.
🎯 What it does: Propose Shifted Autoencoders (SAE), which directly improves the consistency of point annotations by randomly translating point annotations and restoring them to their original positions before training the counting model.
SHINE: Saliency-aware HIerarchical NEgative Ranking for Compositional Temporal Grounding
Zixu Cheng (Queen Mary University of London), Yu Kong (Michigan State University)
CodeRetrievalTransformerLarge Language ModelVision Language ModelVideo
🎯 What it does: Proposes the SHINE method, which generates semantically feasible hard negative samples using large language models and enhances the combinatorial generalization ability of video temporal retrieval through a coarse-to-fine two-level significance ranking.
🎯 What it does: Proposes an audio-visual twin network (AVSiam) that uses a shared vision Transformer to uniformly process audio and visual inputs, and performs self-supervised pre-training through multi-scale random masking, contrastive matching, and reconstruction objectives.
SignGen: End-to-End Sign Language Video Generation with Latent Diffusion
Fan Qi (Tianjin University of Technology), Huaiwen Zhang (Tianjin University of Technology)
CodeGenerationPose EstimationDepth EstimationVision Language ModelDiffusion modelOptical FlowVideoTextMultimodality
🎯 What it does: Developed an end-to-end text-to-sign language video generation system called SignGen based on Latent Diffusion Models (LDMs), directly mapping textual descriptions to complete sign language videos (including body, hand, and facial expressions), eliminating traditional intermediate steps such as gloss or pose prediction.
🎯 What it does: Propose a unified query-based framework called SimPB, which can simultaneously perform Perspective 2D object detection and Bird's-eye view (BEV) 3D object detection under multi-camera inputs;