🎯 What it does: This paper presents the BlurHand dataset and the BlurHandNet model for recovering 3D hand mesh sequences from a single blurred hand image.
🎯 What it does: The deep equilibrium model (DEQ) is transferred to the task of facial keypoint detection, proposing the Landmark DEQ (LDEQ) model, which uses a one-shot root solver to directly solve the heatmap equilibrium points during forward propagation, significantly reducing memory and computational costs; and introduces a temporal consistency constraint through a 'non-recursive recursion' (RwR) mechanism during video inference to suppress keypoint jitter.
🎯 What it does: A regularized vector quantization framework is proposed, integrating deterministic and randomized quantization to achieve high-quality tokenized image synthesis.
🎯 What it does: This paper systematically evaluates the robustness and uncertainty estimation of various modern semantic segmentation models (including ResNet, ConvNeXt, SETR, SegFormer, Segmenter, etc.) under natural domain shifts, exploring the impact of calibration methods (temperature scaling, cluster calibration, local temperature scaling) on calibration, misclassification detection, and OOD detection.
🎯 What it does: This paper proposes a renderable neural radiance map (RNR-Map) that embeds environmental visual information into a grid map using latent codes, achieving image localization and goal-oriented navigation.
🎯 What it does: A visual object tracking-specific representation learning method named Masked Appearance Transfer (MAT) is proposed, which jointly encodes the template and search area using an encoder-decoder architecture, and establishes target correspondence through non-trivial reconstruction objectives.
Reproducible Scaling Laws for Contrastive Language-Image Learning
Mehdi Cherti (Juelich Supercomputing Center Research Center Juelich), Jenia Jitsev (Juelich Supercomputing Center Research Center Juelich)
CodeClassificationRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Systematically conducted large-scale learning experiments on CLIP to explore the impact of model size, data scale, and the number of training samples on downstream task performance, forming publicly reproducible scaling laws.
🎯 What it does: This study proposes a cross-domain learning framework for facial anti-spoofing, SA-FAS, which enhances generalization ability by maintaining domain features that are separable and aligned.
🎯 What it does: Designed and implemented the Federated Prototypes Learning (FPL) scheme, which utilizes clustering prototypes and unbiased prototypes to enhance model generalization and stability in a federated learning environment with domain shift.
🎯 What it does: This paper proposes a new optical flow estimation framework: first pre-training a feature matching extractor using Geometric Image Matching (GIM) data, and then fine-tuning it for the optical flow task, aiming to enhance feature representation and cross-dataset generalization.
🎯 What it does: A framework for OOD detection called MOOD is proposed, which is based on Masked Image Modeling (MIM). It learns the ID distribution features through a self-supervised reconstruction task and then uses the Mahalanobis distance for anomaly detection.
🎯 What it does: View dynamic expression recognition as a weakly supervised multi-instance learning problem, proposing the M3DFEL framework: first, divide the video into 3D instances and use 3D-CNN to learn short-term temporal sequences; then use BiLSTM, attention mechanisms, and dynamic multi-instance normalization to aggregate instances, capturing long-term temporal sequences, ultimately obtaining expression predictions.
🎯 What it does: A contrastive learning framework is proposed that uses finite discrete tokens (FDT) as a shared base to map images and text into a unified discrete space, achieving cross-modal alignment.
Revisiting Prototypical Network for Cross Domain Few-Shot Learning
Fei Zhou (Northwestern Polytechnical University), Yanning Zhang (University of Wollongong)
CodeDomain AdaptationKnowledge DistillationImage
🎯 What it does: This paper proposes a local-global knowledge distillation framework LDP-net to enhance the generalization performance of prototype networks in cross-domain few-shot learning.
Revisiting Rotation Averaging: Uncertainties and Robust Losses
Ganlin Zhang (ETH Zurich), Daniel Barath (ETH Zurich)
CodePose EstimationOptimizationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes to directly use the covariance uncertainty of two-view geometry for rotation averaging in global SfM to improve camera pose estimation.
🎯 What it does: This paper proposes a branch structure called Spatial-Temporal Auxiliary Network (STAN) to transfer image-text pre-trained models like CLIP to video tasks, addressing the issue of simultaneously utilizing high-level semantic knowledge and low-level visual patterns.
Yaqing Ding (Lund University), Kalle Åström (Lund University)
CodePose EstimationImage
🎯 What it does: This paper proposes a P3P solver based on the intersection of two conjugate cones. By analyzing the geometric significance of the roots of a cubic equation, specialized solving strategies are developed for different root configurations, achieving fast and numerically stable camera pose estimation.
🎯 What it does: This paper proposes training the Vision Transformer (ViT) directly on the DCT series of JPEG encoding, without fully decoding to RGB, thereby significantly reducing data loading costs and improving training/inference speed.
RIATIG: Reliable and Imperceptible Adversarial Text-to-Image Generation With Natural Prompts
Han Liu (Washington University in St. Louis), Ning Zhang (Washington University in St. Louis)
CodeGenerationAdversarial AttackLarge Language ModelGenerative Adversarial NetworkImageText
🎯 What it does: A reliable and covert targeted adversarial attack method for text-image generation models, named RIATIG, is proposed, which can generate natural prompts that are semantically aligned with the target image but have almost no relevance to the text content.
🎯 What it does: A detection method based on rigid object visibility guidance is proposed to improve 2D bounding box detection in scenes with severe occlusion for 6D object pose estimation.
🎯 What it does: This paper studies a visual pre-training framework that unifies Masked Image Modeling (MIM) with language supervision (CLIP), proposing masked visual reconstruction in the language semantic space (RILS);
🎯 What it does: A robust adaptive angular margin loss (RMLVQA) is proposed to suppress language bias in visual question answering and achieve better generalization without using data augmentation.
Robust Multiview Point Cloud Registration With Reliable Pose Graph Initialization and History Reweighting
Haiping Wang (Wuhan University), Bisheng Yang (Wuhan University)
CodePose EstimationOptimizationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes a multi-view point cloud registration method that combines learning-based sparse graph construction with historical reweighting iterative least squares (IRLS).
🎯 What it does: A non-local network based on variational Bayesian inference is proposed for filtering outliers in 3D point cloud registration, along with a voting-based inlier search method.
🎯 What it does: A StyleGAN inverse image restoration framework is proposed that does not require separate parameter tuning for each level of denoising or task; robust recovery for various degradation tasks such as upsampling, denoising, artifact removal, and inpainting is achieved through a three-stage gradual expansion of the latent space and the use of conservative normalized gradient descent optimization.
RONO: Robust Discriminative Learning With Noisy Labels for 2D-3D Cross-Modal Retrieval
Yanglin Feng (Sichuan University), Peng Hu (Institute for Infocomm Research A*STAR)
CodeRetrievalContrastive LearningMultimodality
🎯 What it does: A robust learning framework RONO for 2D-3D cross-modal retrieval is proposed, which can effectively learn cross-modal discrimination and invariant representation under the condition of noisy labels.
Rotation-Invariant Transformer for Point Cloud Matching
Hao Yu (TU Munich), Slobodan Ilic (TU Munich)
CodeTransformerPoint Cloud
🎯 What it does: This paper proposes a Transformer model called RoITr that maintains rotation invariance under arbitrary poses for point cloud matching.
🎯 What it does: The study investigated the second-order curvature of the loss for training deep networks with respect to input samples, finding that low-curvature samples are highly correlated with the 'cleanliness' of the data, and based on this, proposed the SLo-Curves core set selection and training method.
🎯 What it does: A resource-adaptive federated learning framework called ScaleFL is proposed, which can dynamically scale the global model in width and depth according to the computational budget of each client and generate multiple exit sub-models, supporting local training and global aggregation under heterogeneous resources.
🎯 What it does: A method called FLIP is proposed to randomly mask image (optional text) blocks during CLIP training to achieve sparse computation and significantly accelerate training.
🎯 What it does: A scanning path prediction model for 360° images, ScanDMM, is proposed, which uses a deep Markov model to model the time-dependent dynamics of visual attention and generate realistic gaze trajectories.
🎯 What it does: Proposes ScarceNet, which performs animal pose estimation using a small number of labeled and unlabeled images through pseudo-labeling, re-labeling with reusable samples, and student-teacher consistency constraints;
🎯 What it does: The SCOOP method is proposed, which combines self-supervised point cloud correspondence learning and runtime flow refinement to estimate scene flow.
🎯 What it does: A new problem is proposed for recovering the 3D geometry of objects within transparent containers, and the ReNeuS method is introduced to solve this problem.
Seeing What You Miss: Vision-Language Pre-Training With Semantic Completion Learning
Yatai Ji (Tsinghua University), Wei Liu (Tencent)
CodeRecognitionRetrievalTransformerVision Language ModelContrastive LearningImageVideoText
🎯 What it does: A semantic completion learning (SCL) task is proposed and implemented to enhance the global alignment capability between images/videos and text through cross-modal information completion.
🎯 What it does: In this study, the authors propose a speech-driven facial generation model called TalkLip, which significantly enhances the readability and lip-sync quality of the generated videos by using lip-reading experts to supervise the generation process.
Self-Positioning Point-Based Transformer for Point Cloud Understanding
Jinyoung Park (Korea University), Hyunwoo J. Kim (Meta Reality Labs)
CodeRecognitionSegmentationTransformerPoint Cloud
🎯 What it does: A Transformer framework named SPoTr is proposed for point cloud understanding, capable of capturing both local and global geometric contexts simultaneously.
🎯 What it does: This study proposes a Scalable Semantic Transfer (SST) training framework for sharing semantic knowledge across multi-label domains, thereby enhancing the performance of human voxel segmentation networks.
🎯 What it does: Proposes the Cleaner Self framework, which uses TSDF-CAD for clean depth training through a teacher network and employs knowledge distillation to guide the student network, thereby improving performance on semantic scene completion tasks with noisy depth inputs.
Semi-Supervised 2D Human Pose Estimation Driven by Position Inconsistency Pseudo Label Correction Module
Linzhi Huang (Beijing University of Posts and Telecommunications), Jieping Ye (Beike)
CodePose EstimationImage
🎯 What it does: A semi-supervised 2D human pose estimation framework (SSPCM) based on a position inconsistency pseudo-label correction module is proposed, combined with pseudo keypoint-aware Cut-Occlude enhancement (SSCO).
Semi-Supervised Domain Adaptation With Source Label Adaptation
Yu-Chu Yu (National Taiwan University), Hsuan-Tien Lin (National Taiwan University)
CodeDomain AdaptationImage
🎯 What it does: This paper proposes a Source Label Adaptation (SLA) framework for Semi-Supervised Domain Adaptation (SSDA), which dynamically corrects source domain labels to better match the target domain feature space, thereby improving classification performance.
SeqTrack: Sequence to Sequence Learning for Visual Object Tracking
Xin Chen (Dalian University of Technology), Han Hu (Microsoft Research)
CodeObject TrackingTransformerVideo
🎯 What it does: A SeqTrack framework is proposed that transforms visual object tracking into a sequence generation task, utilizing a simple encoding-decoding Transformer to complete object bounding box localization.
SFD2: Semantic-Guided Feature Detection and Description
Fei Xue (University of Cambridge), Roberto Cipolla (University of Cambridge)
CodeObject DetectionSegmentationAutonomous DrivingConvolutional Neural NetworkSimultaneous Localization and MappingImage
🎯 What it does: This paper studies how to utilize semantic information to enhance the robustness of feature detection and description in visual localization tasks, and proposes a feature network that implicitly embeds semantics through semantic guidance during the training phase.
SfM-TTR: Using Structure From Motion for Test-Time Refinement of Single-View Depth Networks
Sergio Izquierdo (University of Zaragoza), Javier Civera (University of Zaragoza)
CodeDepth EstimationAutonomous DrivingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes the SfM-TTR method, which utilizes sparse SfM reconstruction as a self-supervised signal during testing for fine-tuning the network in single-view depth estimation.
🎯 What it does: A shadow removal framework called ShadowDiffusion based on diffusion models is proposed, which can recover shadow areas and remove shadows from a single image.
🎯 What it does: This paper proposes a neural field method utilizing shadow ray supervision (ShadowNeuS), which can reconstruct a complete 3D scene from single-view multi-light source shadows or RGB images.
Sharpness-Aware Gradient Matching for Domain Generalization
Pengfei Wang (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
CodeDomain AdaptationImage
🎯 What it does: The Sharpness-Aware Gradient Matching (SAGM) method is proposed, which improves domain generalization ability by simultaneously minimizing empirical risk, perturbation loss, and the gap between the two, allowing the model to converge to low-loss and flat regions.
Shepherding Slots to Objects: Towards Stable and Robust Object-Centric Learning
Jinwoo Kim (Yonsei University), Seon Joo Kim (Yonsei University)
CodeObject DetectionSegmentationImage
🎯 What it does: A single-view image object-centric learning framework called SLASH is proposed, which addresses the bleeding problem caused by background noise by incorporating two modules, ARK (Attention Refinement Kernel) and IPPE (Intermediate Point Prediction and Encoding), based on Slot Attention, achieving more stable and robust object decomposition.
🎯 What it does: The Corgi model is proposed, which utilizes Shifted Diffusion to generate image embeddings in the CLIP multimodal space, enabling text-to-image generation and supporting fully supervised, semi-supervised, and no-language (image-only) training.
🎯 What it does: A Siamese self-supervised pre-training framework for the Transformer detector DETR is proposed, utilizing two-view contrast to achieve pre-training tasks for region detection and semantic distinction.
🎯 What it does: For the task of semantic image editing, a framework called SIEDOB is proposed, which separates the background and foreground objects for individual generation and fusion, significantly improving editing quality in complex scenes.
SIM: Semantic-Aware Instance Mask Generation for Box-Supervised Instance Segmentation
Ruihuang Li (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
CodeObject DetectionSegmentationImage
🎯 What it does: A method for instance mask generation based on semantic prototypes is proposed, which achieves instance segmentation relying solely on box supervision.
🎯 What it does: This paper proposes a metric learning framework for RGB-IR cross-modal group re-identification (G-ReID) called Closest Permutation Matching (CPM), and based on this, designs a weakly supervised Relation-Aware Module (RAM).
🎯 What it does: This paper proposes an online multi-object tracker GHOST based on the Hungarian algorithm, which combines improved re-identification features and a simple linear motion model to achieve robust association performance.
🎯 What it does: The iterative training method of gradient ascent followed by descent in the early layers of the network (SEAL) aims to enhance the model's generalization and transfer performance.
🎯 What it does: This paper proposes a LiDAR semantic segmentation method for single-source domain generalization, DGLSS, which learns domain-invariant representations using sparse consistency and semantic-related consistency constraints, maintaining good performance in both the training source domain and unseen target domain.
🎯 What it does: This paper proposes a single-image reverse inference backdoor trigger method (SmoothInv) that can recover high-success-rate triggers that are highly similar to the original backdoor from a single clean image.
Skinned Motion Retargeting With Residual Perception of Motion Semantics & Geometry
Jiaxu Zhang (Wuhan University), Zhigang Tu (Wuhan University)
CodePose EstimationTransformerVideo
🎯 What it does: A residual structure motion remapping network R2ET is proposed, which can maintain the source action semantics while avoiding the issues of target character crossover and self-collision in a single-step inference.
SLOPER4D: A Scene-Aware Dataset for Global 4D Human Pose Estimation in Urban Environments
Yudi Dai (Xiamen University), Cheng Wang (Xiamen University)
CodePose EstimationSimultaneous Localization and MappingMultimodalityPoint CloudBenchmark
🎯 What it does: We constructed SLOPER4D—a large-scale global 4D human pose dataset in urban environments that includes LiDAR, camera, and IMU data. Based on this, we conducted benchmark experiments on 3D human pose estimation (HPE) and global human pose estimation (GHPE).
🎯 What it does: A time delay attack on the LiDAR detection pipeline named SlowLiDAR is proposed, which can significantly increase the running time of detection models by perturbing the point cloud or adding points, while maintaining the imperceptibility of the adversarial samples.
SmallCap: Lightweight Image Captioning Prompted With Retrieval Augmentation
Rita Ramos (Instituto Superior Técnico), Yova Kementchedjhieva
CodeGenerationRetrievalDomain AdaptationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageVideoTextRetrieval-Augmented Generation
🎯 What it does: A lightweight image description model called SMALLCAP is designed, which utilizes retrieval-enhanced prompts to generate descriptions and can achieve cross-domain transfer without additional training.
🎯 What it does: A continuous convolution based on self-moving point representation (SMPConv) is proposed, where the convolution kernel parameters are represented by learnable point coordinates, weights, and radii, allowing for the construction of convolution kernels of arbitrary resolution without the need for neural networks.
🎯 What it does: This paper proposes a Soft Augmentation technique that dynamically softens target labels or sample weights based on the degree of transformation, allowing the network to train under more aggressive augmentations without losing performance.
🎯 What it does: A lightweight Soft-Landing (SoLa) module is proposed, placed between the frozen snippet encoder and the TAL head, and trained using self-supervised Similarity Matching to enhance the temporal sensitivity of segment features and alleviate the task gap issue.
🎯 What it does: This paper proposes a method that combines a pre-trained two-dimensional diffusion model with traditional TV-based models to address the problem of three-dimensional medical image reconstruction.
🎯 What it does: This paper proposes a theoretical perspective to explain and solve the oscillation problem in post-training quantization (PTQ), further improving the accuracy of PTQ.
Solving Relaxations of MAP-MRF Problems: Combinatorial In-Face Frank-Wolfe Directions
Vladimir Kolmogorov (Institute of Science and Technology Austria)
CodeOptimizationGraphBenchmark
🎯 What it does: An efficient implementation of LP relaxation for the MAP-MRF problem is proposed using the Frank-Wolfe algorithm combined with in-face directions.
🎯 What it does: Train an audio-to-visual alignment model to generate high-quality scene images corresponding to input audio without labeled or language supervision.
🎯 What it does: A passive data cross-domain gaze estimation method called UnReGA is proposed, which achieves adaptation by reducing sample and model uncertainty.
🎯 What it does: A source-agnostic video domain adaptation method based on spatio-temporal-historical consistency learning (STHC) is proposed, achieving model adaptation by randomly applying spatial and temporal augmentations on the target video and enforcing three types of consistency losses.
🎯 What it does: This paper proposes a sparse annotation semantic segmentation framework based on an Adaptive Gaussian Mixture Model (AGMM), using labeled pixels as the centers of the Gaussian mixtures to generate soft GMM predictions online for self-supervision.
🎯 What it does: This paper studies the application of source-free domain adaptation (SFDA) in video semantic segmentation (VSS) and proposes a spatiotemporal fusion method STPL based on pixel-level contrastive learning.
🎯 What it does: This paper proposes a LiDAR point cloud semantic segmentation method based on spatiotemporal self-supervised learning, utilizing unsupervised clustering and tracking to construct positive samples for end-to-end self-supervised pre-training.
🎯 What it does: A fine-tuning framework named Specialist Diffusion is proposed, which learns unknown styles using a very small number (≤10 images) of images, enabling a pre-trained text-to-image diffusion model to generate images of any object in that style.
🎯 What it does: A neighborhood correlation-aware (NeCA) sRGB noise synthesis framework is proposed, explicitly modeling the signal correlation and neighborhood correlation of noise, generating realistic camera noise directly in the sRGB domain.
🎯 What it does: An end-to-end multi-camera 3D multi-object tracking framework called PF-Track is designed, which integrates past and future spatiotemporal reasoning to achieve high-precision trajectory tracking.
🎯 What it does: A NeRF acceleration framework called SteerNeRF is proposed, which achieves real-time rendering by utilizing low-resolution volumetric rendering and high-resolution 2D neural rendering.
Structure Aggregation for Cross-Spectral Stereo Image Guided Denoising
Zehua Sheng (Zhejiang University), Huaqi Zhang (vivo Mobile Communication Company Ltd.)
CodeRestorationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a structure aggregation-based cross-spectral stereo image guided denoising network (SANet) that can recover details of noisy images from unaligned RGB-NIR image pairs.
🎯 What it does: A structured sparse learning (SSL) framework is proposed, which achieves compression and acceleration of video super-resolution models through pruning residual blocks, recurrent networks, and pixel rearrangement.
🎯 What it does: A cross-domain few-shot learning method based on Style Adversarial Training (StyleAdv) is proposed, aimed at alleviating domain gaps by generating virtual and challenging samples in the style space.
🎯 What it does: A semantic benchmark for identity preservation based on StyleGAN, called StyleIPSB, is proposed, and a three-stage high-fidelity face swapping framework is constructed to achieve precise fusion of the source face identity and target facial attributes.
🎯 What it does: A single-stage image inversion framework called StyleRes is proposed, which learns residual features and performs transformations in a higher-order latent space, achieving high-fidelity reconstruction and high-quality attribute editing.
🎯 What it does: The StyLess attack method is proposed, which generates various stylized models by inserting instance normalization layers into the surrogate model, and jointly uses the gradients of these models to reduce reliance on non-robust style features, significantly enhancing the transferability of black-box attacks.
🎯 What it does: By automatically learning multi-layer super-class graphs and combining them with message passing in graph neural networks, the original features in long-tail visual recognition are corrected and enhanced, thereby improving the recognition performance of tail classes.
🎯 What it does: A supervised mask knowledge distillation (SMKD) framework is proposed to enhance model generalization performance in few-shot learning tasks using Vision Transformer.
🎯 What it does: An adaptive network with system state awareness (SAN) is proposed for achieving low-latency online video understanding on devices with fluctuating computational resources.
🎯 What it does: A single-model self-ensemble adversarial patch attack (T-SEA) is designed to enhance the transferability of adversarial patches in multi-object detectors through strategies such as data self-ensemble, model ShakeDrop, and patch cutout.