🎯 What it does: This paper proposes MaxLin, a framework for reachability verification of CNNs with MaxPool layers, utilizing compact linear approximations to achieve stronger robustness lower bounds.
🎯 What it does: A three-stage universal tumor synthesis framework named DiffTumor is proposed and implemented, which generates early tumors in multiple organs from a small number of annotated tumor samples in CT images, and enhances the performance of tumor detection/separation models through these synthetic data.
🎯 What it does: Proposes a Rolling Mixed Bit (RMB) reading mechanism and RMB-Net network for reconstructing HDR and high frame rate video from single-bit and multi-bit time-varying pulses.
🎯 What it does: A Point Prompt Training (PPT) framework is proposed to achieve collaborative training of 3D representation learning across multiple datasets, avoiding negative transfer.
Towards Learning a Generalist Model for Embodied Navigation
Duo Zheng (Chinese University of Hong Kong), Liwei Wang (Chinese University of Hong Kong)
CodeRobotic IntelligenceTransformerLarge Language ModelPrompt EngineeringTextMultimodality
🎯 What it does: This paper presents NaviLLM, a general-purpose sensory navigation model based on large language models, which unifies various tasks into generative questions through schema-based instruction.
Chen Chen (University of Sydney), Chang Xu (University of Sydney)
CodeGenerationData SynthesisSafty and PrivacyDiffusion modelImage
🎯 What it does: This paper proposes an Anti-Memory Guidance (AMG) framework that combines three targeted guidance strategies to eliminate the memory phenomenon of pre-trained diffusion models without compromising image quality.
🎯 What it does: This paper proposes a new paradigm for constrained image tampering localization (CIML) called CAAA, and based on this paradigm, constructs a large-scale, high-quality MIML dataset, further designing the APSC-Net model for image tampering localization.
🎯 What it does: This paper presents TimeTuner, a plug-in method that reduces truncation errors in the acceleration process of diffusion models by optimizing the time step size at each step.
🎯 What it does: A multi-frequency representation (MFR) network is proposed for image warping, achieving coarse-to-fine image reconstruction by learning features from different frequency bands layer by layer.
🎯 What it does: This paper proposes a 3D point cloud-based anomaly detection and localization framework called IMRNet, and constructs a scalable synthetic dataset named Anomaly-ShapeNet.
Towards the Uncharted: Density-Descending Feature Perturbation for Semi-supervised Semantic Segmentation
Xiaoyang Wang (Xidian University), Jimin Xiao (Beijing Jiaotong University)
CodeSegmentationFlow-based ModelImage
🎯 What it does: This paper proposes a disturbance strategy based on low-density partitioning in feature space called Density-Descending Feature Perturbation (DDFP), which achieves stronger discriminative consistency learning by perturbing features in low-density directions within a semi-supervised semantic segmentation framework.
🎯 What it does: A high-density dataset TSP6K aimed at traffic monitoring scenarios is proposed, and various scene parsing, instance segmentation, and unsupervised domain adaptation methods are evaluated on it. Furthermore, a detail refining decoder is designed to improve segmentation performance in monitoring scenarios.
🎯 What it does: A GAN super-resolution training framework based on wavelet domain loss is proposed, utilizing adversarial and reconstruction losses from SWT subbands to suppress high-frequency artifacts and enhance detail fidelity.
🎯 What it does: A general medical image segmentation framework called Hermes is proposed, which can perform multi-position, multi-modal, and partially labeled multi-task segmentation within a single model.
Zhengqi Xu (Zhejiang University), Jie Song (Zhejiang University)
CodeConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposes the MuDSC (Merging under Dual-Space Constraints) framework, which utilizes an untrained pre-trained model to achieve multi-task model merging through unit matching.
🎯 What it does: This paper proposes a deepfake detection method based on latent space augmentation, called LSDA, aimed at enhancing the model's generalization ability by expanding the forgery space.
Transcending the Limit of Local Window: Advanced Super-Resolution Transformer with Adaptive Token Dictionary
Leheng Zhang (University of Electronic Science and Technology of China), Shuhang Gu (University of Electronic Science and Technology of China)
CodeRestorationSuper ResolutionTransformerImage
🎯 What it does: Proposes the Adaptive Token Dictionary (ATD) Transformer, which utilizes a learnable Token Dictionary to integrate external priors and global information through cross-attention and self-attention, in order to enhance the super-resolution quality of single images.
🎯 What it does: This paper proposes a transferable sparse subnetwork and an efficient fine-tuning method based on weight spectrum, achieving a significant reduction in model size and training costs for Open-Vocabulary Segmentation (OVS).
🎯 What it does: This paper proposes a bionic focus attention (Aggregated Attention) and Convolutional GLU, integrating them into a new visual backbone network called TransNeXt, aimed at addressing the degradation problem of traditional Vision Transformers in deep information mixing.
Troika: Multi-Path Cross-Modal Traction for Compositional Zero-Shot Learning
Siteng Huang (Zhejiang University), Donglin Wang (Westlake University)
CodeClassificationRecognitionTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: This paper addresses the problem of synthetic zero-shot learning and proposes a multi-path cross-modal coupling model called Troika, which explicitly models three types of semantics: state, object, and combination through three branches.
🎯 What it does: A testing time adaptation framework for event camera video frame interpolation (TTA‑EVF) is proposed, which can adapt the network online using only low frame rate videos and event streams in the target domain.
Tumor Micro-environment Interactions Guided Graph Learning for Survival Analysis of Human Cancers from Whole-slide Pathological Images
Wei Shao (Nanjing University of Aeronautics and Astronautics), Peng Wan (Nanjing University of Aeronautics and Astronautics)
CodeClassificationSegmentationExplainability and InterpretabilityGraph Neural NetworkImageBiomedical Data
🎯 What it does: Developed a TMEGL model based on whole-slide images to predict the survival of human cancer patients using tumor microenvironment (TME) interactions.
🎯 What it does: This paper presents the TUMTraf V2X multimodal V2X cooperative perception dataset for real traffic scenarios and develops the CoopDet3D cooperative 3D object detection model based on it.
🎯 What it does: This paper proposes a differentiable visual prompt method (Tune-An-Ellipse) that utilizes the visual prompting capability of CLIP to achieve zero-shot referential expression localization by iteratively optimizing the parameters of an ellipse.
🎯 What it does: During the fine-tuning process of pre-trained models, it was found that existing methods failed to effectively reduce model complexity. The TSRS (Tuning Stable Rank Shrinkage) regularization is proposed, which utilizes noise sensitivity to constrain the model's stable rank, thereby reducing structural risk and enhancing generalization.
U-VAP: User-specified Visual Appearance Personalization via Decoupled Self Augmentation
You Wu (Institute of Computing Technology, Chinese Academy of Sciences), Jintao Li (Institute of Computing Technology, Chinese Academy of Sciences)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringDiffusion modelImageText
🎯 What it does: Under limited reference images, users can specify visual attributes through text descriptions to achieve refined visual appearance personalization.
🎯 What it does: Designed and implemented the ULIP-2 framework, which automatically generates fine-grained language descriptions of 3D shapes through large multimodal models, constructs an unannotated tri-modal (point cloud, image, text) dataset, and conducts pre-training to enhance 3D representation learning effectiveness.
UltrAvatar: A Realistic Animatable 3D Avatar Diffusion Model with Authenticity Guided Textures
Mingyuan Zhou (OPPO US Research Center), Guojun Qi (Westlake University)
CodeGenerationDiffusion modelImageMesh
🎯 What it does: UltrAvatar has been developed to generate animatable 3D avatars with realistic lighting editable PBR textures from text prompts or a single facial image.
Uncertainty Visualization via Low-Dimensional Posterior Projections
Omer Yair (Technion Israel Institute of Technology), Tomer Michaeli (Technion Israel Institute of Technology)
CodeImage TranslationRestorationGenerationDiffusion modelImageBiomedical Data
🎯 What it does: A framework is proposed for estimating and visualizing posterior distributions in low-dimensional subspaces for image inversion problems.
🎯 What it does: This paper analyzes the role of discriminative and diversity loss in source-free domain adaptation (SFDA) from a theoretical perspective and proposes an improved SFDA method based on this analysis.
Unexplored Faces of Robustness and Out-of-Distribution: Covariate Shifts in Environment and Sensor Domains
Eunsu Baek (Seoul National University), Hyung-Sin Kim (Seoul National University)
CodeDomain AdaptationAnomaly DetectionImage
🎯 What it does: This paper constructs a controllable testing platform, ES-Studio, using real cameras to capture 202k images under different lighting conditions and camera sensor parameters, creating the ImageNet-ES dataset, and conducting experiments on OOD detection, domain generalization, and camera sensor control on this dataset.
Unified Entropy Optimization for Open-Set Test-Time Adaptation
Zhengqing Gao (Institute of Automation, Chinese Academy of Sciences), Cheng-Lin Liu (Institute of Automation, Chinese Academy of Sciences)
CodeDomain AdaptationOptimizationImage
🎯 What it does: A unified entropy optimization framework (UniEnt/UniEnt+) is proposed for adaptive and unknown category detection of pre-trained models in an open-set testing environment (Open-Set TTA).
Unifying Top-down and Bottom-up Scanpath Prediction Using Transformers
Zhibo Yang (Stony Brook University), Dimitris Samaras (Stony Brook University)
CodeTransformerImage
🎯 What it does: A unified Human Attention Transformer (HAT) model is proposed to simultaneously predict goal-directed (top-down) and free viewing (bottom-up) scan paths;
🎯 What it does: A new parameter-efficient transfer learning method called UniPT is proposed, which achieves transfer without the need for backpropagation through the main network by adding a lightweight parallel network alongside the pre-trained model.
🎯 What it does: A universal large-kernel convolutional network called UniRepLKNet is proposed, which can achieve efficient recognition, segmentation, and detection in image tasks, as well as perform excellently in multimodal tasks such as audio, video, point clouds, and time series.
🎯 What it does: A unified video segmentation framework called UniVS is proposed, which utilizes prompts (visual or textual) as queries to uniformly handle all category-specific and prompt-specific video segmentation tasks.
🎯 What it does: We propose H-SAM, a prompt-free version of the Segment Anything Model, which achieves efficient fine-tuning and fine-grained segmentation of medical images through a two-stage hierarchical decoder.
🎯 What it does: This paper proposes a completely unsupervised and semi-supervised cross-view geographic localization framework that utilizes unlabeled data to achieve retrieval from ground images to satellite images.
🎯 What it does: This paper proposes SGPT, a visual prompt tuning framework that combines shared prompts and grouped prompts in federated learning, enabling the global model to adapt to the data distribution of different clients without local fine-tuning.
🎯 What it does: This paper proposes a cost-effective visual post-processing (VPO) dataset construction scheme and a supervised contrastive learning-based audio-visual segmentation method (CAVP) to improve the audio-visual segmentation task.
🎯 What it does: A high-transferable, prompt-free adversarial attack method UAD is proposed for prompt-based segmentation models (such as SAM), which can significantly disrupt segmentation results across various models and prompts.
🎯 What it does: A neural implicit representation based on the unsigned orthogonal distance field (UODF) is proposed for the precise reconstruction of various 3D shapes.
🎯 What it does: A unified unsupervised image segmentation framework U2Seg is proposed, capable of simultaneously performing instance segmentation, semantic segmentation, and panoptic segmentation tasks.
Unveiling Parts Beyond Objects: Towards Finer-Granularity Referring Expression Segmentation
Wenxuan Wang (Institute of Automation, Chinese Academy of Sciences), Jing Liu (Institute of Automation, Chinese Academy of Sciences)
CodeObject DetectionSegmentationTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This paper proposes a Multi-Granularity Referring Expression Segmentation task (MRES), constructs the RefCOCOm evaluation benchmark, releases the largest-scale 32M visual localization dataset, and introduces a unified UniRES model to achieve object and part-level referring segmentation.
🎯 What it does: A deep fusion audio-visual Transformer is proposed, utilizing learnable fusion tokens to achieve early fusion through dense local interactions, and is pretrained under a self-supervised masked reconstruction framework.
Unveiling the Unknown: Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation
Fuli Wan (Xidian University), Cheng Deng (Xidian University)
CodeDomain AdaptationContrastive LearningImage
🎯 What it does: This paper proposes an open-set source unsupervised domain adaptation framework in which source data is unavailable and the target domain contains unknown classes. It utilizes an unknown diffuser to actively mine unknown classes in the target domain from a source pre-trained model, thereby achieving a bidirectional enhancement of knowledge transfer for known classes and generalization for unknown classes.
🎯 What it does: The first large-scale high-resolution underwater video enhancement benchmark UVEB has been constructed, and the first supervised underwater video enhancement network UVE-Net has been proposed.
🎯 What it does: This paper proposes an online 'Virtual Assurance Amplification Attack' (VA3), which significantly increases the probability of generating infringing images by text-to-image generation models under probabilistic copyright protection through multiple interactions, and provides an adversarial prompt optimization algorithm called Anti-NAF for NAF protection.
Validating Privacy-Preserving Face Recognition under a Minimum Assumption
Hui Zhang (Anhui University), Xuejun Li (Anhui University)
CodeRecognitionSafty and PrivacyGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a privacy verification method called Map V based on minimal assumptions (1k1c), which utilizes deep image priors and zero-order gradient estimation to attack privacy-protecting facial recognition systems with only a limited number of queries.
VCoder: Versatile Vision Encoders for Multimodal Large Language Models
Jitesh Jain (SHI Labs), Humphrey Shi (Picsart AI Research)
CodeObject DetectionSegmentationDepth EstimationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: To address the shortcomings of multimodal large language models (MLLM) in object perception and counting tasks, the VCoder controller is proposed, and the COCO Segmentation Text (COST) dataset is constructed for training and evaluating the object recognition, counting, and depth order perception capabilities of MLLMs.
🎯 What it does: Without the need for additional training, vid-TLDR is proposed to perform early token merging on video Transformers to reduce computational costs and improve performance.
🎯 What it does: A Video Triplet Transformer (VTT) framework is proposed for video and video tasks (such as video fusion, video enhancement, and video de-moiré) to achieve visual consistency and temporal consistency in videos by modeling three types of spatiotemporal variation patterns (short-term spatial, long-term global, and long-term dynamic).
🎯 What it does: This paper proposes a regression-based multi-frame human pose estimation framework called DSTA, which can directly regress keypoint coordinates from video frames.
CodeRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a new visual language model (VLM) pre-training scheme—VILA, systematically exploring the key designs for visual language pre-training on LLMs, and achieving stronger multimodal reasoning and instruction-following capabilities based on this.
🎯 What it does: A weakly supervised learning-based unpaired virtual immunohistochemistry (IHC) staining framework (Confusion-GAN) is proposed, capable of generating high-quality, pathologically consistent IHC images without the need for H&E-IHC aligned images.
🎯 What it does: DINOv is proposed, a unified visual context prompting framework that can simultaneously perform referential segmentation and general segmentation tasks.
🎯 What it does: A visual point cloud prediction pre-training framework called ViDAR is proposed, using future point cloud prediction as a pre-training task to enhance the performance of vision-driven perception, prediction, and planning.
🎯 What it does: This paper proposes VIT-LENS, an omni-modal representation learning framework that utilizes a pre-trained ViT to project any new modality into the visual space through a Lens module and align it with foundational models like CLIP.
🎯 What it does: This study proposes an orthogonal projection-based knowledge distillation method that can directly project features while ensuring the invariance of intra-batch feature similarity, thereby maximizing knowledge transfer, applicable to classification, detection, and generation tasks.
🎯 What it does: This paper proposes the VoCo framework for self-supervised pre-training of 3D medical images, predicting the contextual position of sub-volumes based on volume contrast learning.
🎯 What it does: This paper proposes a voxelized 3D environment representation (VER) by projecting multi-view 2D features into a 3D voxel grid, and performing coarse-to-fine feature extraction and multi-task learning in this space to achieve joint predictions of 3D occupancy, room layout, and 3D detection, thereby providing a more complete scene understanding for Visual Language Navigation (VLN);
🎯 What it does: This paper proposes a Visual Reference Prompt (VRP) encoder, which is integrated with the Segment Anything Model (SAM) to form VRP-SAM, thereby supporting semantic segmentation of target images directly using annotated reference images (points, boxes, lines, masks).
VS: Reconstructing Clothed 3D Human from Single Image via Vertex Shift
Leyuan Liu (National Engineering Research Center for E-Learning Central China Normal University), Jingying Chen (National Engineering Research Center for E-Learning Central China Normal University)
🎯 What it does: A method for 3D human body reconstruction from a single image based on a two-stage vertex shift has been proposed, which achieves high-fidelity and defect-free 3D reconstruction of humans wearing loose clothing while maintaining the structure of the human body.
🎯 What it does: A general visual salient object and camouflage object detection framework, VSCode, is proposed, capable of handling multi-modal SOD and COD tasks in one go.
VSRD: Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection
Zihua Liu (Tokyo Institute of Technology), Masatoshi Okutomi (Tokyo Institute of Technology)
CodeObject DetectionAutonomous DrivingImage
🎯 What it does: A weakly supervised 3D object detection framework VSRD based on multi-view automatic labeling is proposed, which optimizes 3D bounding boxes using 2D instance masks and generates pseudo-labels to train a monocular 3D detector.
Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration
Chen Zhao (Nanjing Normal University), Chengwei Hu (Nanjing Normal University)
CodeRestorationTransformerDiffusion modelImage
🎯 What it does: This paper proposes a seabed image enhancement framework WF-Diff based on frequency domain features and diffusion models, which can first repair color distortion in the frequency domain and then refine details using a frequency domain residual diffusion model.
Weak-to-Strong 3D Object Detection with X-Ray Distillation
Alexander Gambashidze (Artificial Intelligence Research Institute), Ilya Makarov (Artificial Intelligence Research Institute)
CodeObject DetectionAutonomous DrivingKnowledge DistillationPoint CloudTime Series
🎯 What it does: Proposes the X-Ray Teacher framework, which enhances 3D detection performance in sparse and occluded scenes by utilizing LiDAR time series to generate Object-Complete Frames and employing Teacher-Student knowledge distillation.
Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle
Hyeokjun Kweon (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
CodeSegmentationPoint Cloud
🎯 What it does: A weakly supervised point cloud semantic segmentation framework (REAL) is proposed, which combines artificial prior label generation based on SAM (Segment Anything Model) with active learning.
🎯 What it does: A Continuous Testing Adaptation (CTA) method for object detection is proposed, achieving online adaptation in a constantly changing testing domain.
Why Not Use Your Textbook? Knowledge-Enhanced Procedure Planning of Instructional Videos
Kumaranage Ravindu Yasas Nagasinghe (Mohamed bin Zayed University of Artificial Intelligence), Muhammad Haris Khan (Mohamed bin Zayed University of Artificial Intelligence)
CodeDiffusion modelVideo
🎯 What it does: This paper proposes a knowledge-enhanced program planning framework called KEPP, which utilizes a probabilistic program knowledge graph generated from the training set to assist in planning action sequences from the initial visual state to the target visual state.
WildlifeMapper: Aerial Image Analysis for Multi-Species Detection and Identification
Satish Kumar (University of California Santa Barbara), B.S. Manjunath (University of California Santa Barbara)
CodeRecognitionObject DetectionTransformerImage
🎯 What it does: WildlifeMapper (WM) has been developed, an end-to-end model based on Transformer for detecting, locating, and identifying multiple wildlife species in high-resolution aerial images. A large-scale annotated dataset of the Masai Mara ecosystem has been publicly released, containing 21 species and 28k target boxes.
WOUAF: Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models
Changhoon Kim (Arizona State University), Yezhou Yang (Arizona State University)
CodeGenerationData SynthesisSafty and PrivacyConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: A user fingerprint embedding method based on weight modulation, WOUAF, is proposed, which can quickly generate user-specific fingerprints on the Stable Diffusion decoder, enabling user attribution for the model distributor.
CodeExplainability and InterpretabilityConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: A unified framework called WWW is proposed, which can simultaneously explain the three layers of meaning of a model: 'what', 'where', and 'why'.
🎯 What it does: This study proposes X-3D, which significantly improves point cloud classification, segmentation, and detection performance by explicitly constructing local geometric structures in the original space and generating shared dynamic structural kernels based on this.
🎯 What it does: This study investigates action recognition across datasets in first-person perspective videos and proposes the X-MIC framework for cross-modal instance conditioning in a frozen CLIP embedding space.
CodeObject DetectionConvolutional Neural NetworkPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Developed YOLO-World, a real-time open vocabulary object detection framework that combines Vision-Language pre-training with the re-parameterized RepVL-PAN to achieve zero-shot open vocabulary detection.
🎯 What it does: This paper proposes YolOOD, which utilizes the 'objectness' and classification scores of the YOLO object detection model to achieve out-of-distribution (OOD) detection for multi-label images.
🎯 What it does: An adaptive teacher-student collaboration method is proposed, where a few-step student model is first used to generate images, and then a threshold is applied to decide whether to use the teacher model for further optimization.
🎯 What it does: The ZePT framework is proposed, which can perform zero-shot tumor segmentation using only annotated organ data, and simultaneously segment known organs and unknown tumors during inference.
🎯 What it does: A training-free, layout-based text-to-image generation framework called Zero-Painter is proposed, which can generate images that conform to the shape and text attributes from object masks, corresponding text descriptions, and global prompts.
🎯 What it does: This paper proposes a zero-shot structure-preserving diffusion model that guides the tone mapping from HDR to LDR using structural information.