IEEE/CVF International Conference on Computer Vision Β· 833 papers
ReCamMaster: Camera-Controlled Generative Rendering from A Single Video
Jianhong Bai (Kuaishou Tech), Di Zhang (Kuaishou Tech)
CodeGenerationData SynthesisDiffusion modelVideo
π― What it does: ReCamMaster is a camera control video re-rendering framework based on a pre-trained text-to-video diffusion model, capable of generating new videos that conform to specified camera trajectories from a single input video.
π― What it does: In the video object segmentation task, a ReferDINO model based on GroundingDINO is proposed, achieving end-to-end text description-driven video object segmentation.
π― What it does: For the task of referring expression comprehension (REC) for extremely small objects, the authors constructed a new SOREC dataset and proposed the ProgressiveβIterative Zooming Adapter (PIZA) module, enabling the model to progressively zoom in and locate the target in an autoregressive manner.
Qing Jiang (South China University of Technology), Lei Zhang (International Digital Economy Academy)
CodeRecognitionObject DetectionRetrievalTransformerLarge Language ModelImageMultimodalityRetrieval-Augmented Generation
π― What it does: Proposed the human multi-instance referencing task and its corresponding dataset HumanRef, along with the retrieval-based large language model RexSeek.
Reflect-DiT: Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection
Shufan Li (University of California Los Angeles), Aditya Grover (University of California Los Angeles)
CodeObject DetectionGenerationData SynthesisTransformerVision Language ModelDiffusion modelImageText
π― What it does: Introducing a reflection mechanism during inference in text-to-image diffusion models allows the model to gradually improve generation results by utilizing previous images and natural language feedback during the generation process.
π― What it does: This paper proposes a Region-level Clustering Discrimination (RICE) framework, constructing a billion-level candidate region dataset, and incorporates a Region Transformer layer and a unified region clustering loss into the visual encoder to enhance the visual model's ability to perceive objects and OCR regions.
π― What it does: This paper proposes a region-level data attribution framework for text-to-image diffusion models, capable of locating the most influential image regions in training samples and providing attribution scores.
Hongchi Ma (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
CodeAnomaly DetectionTransformerVision Language ModelImageMultimodality
π― What it does: This paper proposes a retrieval-enhanced multimodal prompt fusion framework, ReMP-AD, for industrial visual anomaly detection with few samples, which includes two main modules: Intra-class Token Retrieval (ICTR) and Visual-Language Prior Fusion (VLPF).
RePoseD: Efficient Relative Pose Estimation With Known Depth Information
Yaqing Ding (Comenius University in Bratislava), Zuzana Kukelova (Czech Technical University in Prague)
CodePose EstimationDepth EstimationComputational EfficiencySimultaneous Localization and MappingImage
π― What it does: The research utilizes monocular depth estimation (MDE) results to propose a new relative pose solver for cameras that can simultaneously estimate scale/offset parameters and relative pose.
π― What it does: Proposes the COD-VAE two-stage autoencoder framework, which compresses 3D shapes into only 64 1D latent vectors, significantly improving generation efficiency and quality.
π― What it does: Proposes the Resonance (Re) model, which predicts pedestrian trajectories using the concepts of vibration and resonance, breaking down trajectory randomness into spontaneous vibration and social vibration, generated through noise sampling respectively;
Rethinking Cross-Modal Interaction in Multimodal Diffusion Transformers
Zhengyao Lv (University of Hong Kong), Kwan-Yee K. Wong (University of Hong Kong)
CodeGenerationData SynthesisTransformerSupervised Fine-TuningVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
π― What it does: A parameter-efficient Temperature Adaptive Cross-modal Attention (TACA) method is proposed to rebalance cross-modal attention in multi-modal diffusion transformers, thereby enhancing text-image alignment.
π― What it does: Proposes the Revised Direct Preference Optimization (RDPO) method, which aligns text-to-image diffusion models through human preferences and addresses the risks of reward estimation bias in intermediate steps and the increased probability of non-preferred samples in traditional DPO.
π― What it does: A set of unlearning techniques utilizing the CLIP model is proposed to construct a true inductive few-shot benchmark, and a self-enhanced prompt tuning method (SEPRES) is introduced based on this benchmark.
π― What it does: In response to the fusion degradation phenomenon in RGB-IR multimodal object detection, this paper identifies through linear probe evaluation that multimodal joint training leads to insufficient single-modal feature learning. Consequently, it proposes the Mβ―D-LIF framework: Mono-Modality Distillation (Mβ―D) enhances the capability of single-modal encoders, while Local Illumination-Aware Fusion (LIF) dynamically weights the fusion of RGB and IR features through brightness prediction.
π― What it does: A light field super-resolution decoder based on implicit image functions (SEIIF) is proposed, achieving arbitrary scale light field super-resolution through two upsampling modes: spatial (SIIF) and epipolar (EIIF).
π― What it does: A deep convolutional reverse convolution operation is proposed, along with the construction of a corresponding network module, applied to image denoising, super-resolution, and deblurring tasks.
π― What it does: The first benchmark for adversarial patch defense against object detection (APDE) is proposed, systematically evaluating 11 defense methods, 13 patch attacks, and 11 detectors, and constructing a multi-type patch dataset of 94,000 images.
π― What it does: This paper proposes a new dual-branch offset learning paradigm to simultaneously adjust image features and class representations, thereby addressing the feature-class misalignment issue in traditional pixel-level classification methods for lightweight semantic segmentation.
π― What it does: This paper presents research on point cloud completion for real-world applications, constructing the RealPC dataset, which contains approximately 40,000 pairs and 21 categories of industrial structures, and explores the role of 0-dimensional persistent homology (PH) in point cloud completion.
π― What it does: This paper re-examines pool-based prompt learning in few-shot class-incremental learning (FSCIL) and finds that traditional stacking prompts in the token dimension leads to performance drops for new classes due to information saturation. It proposes LGSP-Prompt, which reconstructs prompts in the spatial dimension through local and global spatial prompts to address the saturation issue.
π― What it does: A reward-guided Gaussian Splatting extended reconstruction framework (RGE-GS) is proposed, which can utilize prior images generated by diffusion models to complete missing road scenes in a single scan.
RobAVA: A Large-scale Dataset and Baseline Towards Video based Robotic Arm Action Understanding
Baoli Sun (Dalian University of Technology), Zhiyong Wang (Dalian University of Technology)
CodeRecognitionAnomaly DetectionRobotic IntelligenceTransformerLarge Language ModelContrastive LearningVideo
π― What it does: A large-scale robot arm action video dataset called RobAVA is proposed, along with a motion recognition model based on atomic attributes, AGPT-Net.
RoboTron-Mani: All-in-One Multimodal Large Model for Robotic Manipulation
Feng Yan (Meituan), Lin Ma (Meituan)
CodeRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningImageTextMultimodality
π― What it does: RoboTron-Mani is proposed, a multimodal large model capable of accepting multi-view images, camera parameters, and text instructions, and outputting actions, images, and occupancy maps, while also constructing a unified RoboData dataset.
π― What it does: A robust multi-view learning framework RML is proposed, which utilizes sample-level attention to fuse multi-view representations and achieves representation alignment and robustness enhancement through contrastive learning with simulated noise and unavailable perturbations.
π― What it does: This paper proposes an adaptive framework for single image denoising based on deep Gaussian priors, where a Gaussian denoising model is first pre-trained using a self-supervised approach. Then, for each image to be processed, a pixel pool is constructed, and pseudo-instances are randomly sampled for a small number of iterative fine-tuning, allowing the model to quickly adapt to unknown noise distributions.
π― What it does: Developed the RoCo-Sim simulation framework for generating multi-view consistent foreground insertions and realistic training data from a single fixed-view image.
RTMap: Real-Time Recursive Mapping with Change Detection and Localization
Yuheng Du (CaiNiao Inc.), Qiang Li (CaiNiao Inc.)
CodeAutonomous DrivingOptimizationTransformerSimultaneous Localization and MappingPoint Cloud
π― What it does: RTMap is proposed, an end-to-end real-time online high-definition map construction framework that integrates map localization, structural change detection, and multi-vehicle collaborative updates.
S3E: Self-Supervised State Estimation for Radar-Inertial System
Shengpeng Wang (Huazhong University of Science and Technology), Wei Wang (Wuhan University)
CodePose EstimationAutonomous DrivingConvolutional Neural NetworkSimultaneous Localization and MappingPoint Cloud
π― What it does: A self-supervised S3E radar-inertial system state estimation method is proposed, which fuses the amplitude spectrum of millimeter-wave radar with IMU data to directly extract geometrically consistent feature points from the radar spectrum and estimate the vehicle's instantaneous speed.
π― What it does: A street view reconstruction framework S3R-GS based on 3D Gaussian Splatting is proposed, significantly reducing the computational load and time for large-scale street view reconstruction.
Safeguarding Vision-Language Models: Mitigating Vulnerabilities to Gaussian Noise in Perturbation-based Attacks
Jiawei Wang (University of Science and Technology of China), Kin-Man Lam (Hong Kong Polytechnic University)
CodeAdversarial AttackTransformerVision Language ModelDiffusion modelImageTextMultimodality
π― What it does: This paper proposes the Robust-VLGuard training scheme by constructing the Robust-VLGuard visual-text security dataset and performing noise augmentation with LoRA fine-tuning. It then designs the DiffPure-VLM defense pipeline by combining the noise distribution transformation of diffusion models, both of which enhance the robustness and security of VLM against Gaussian noise and optimized visual perturbations.
π― What it does: Using monocular RGB video combined with shape templates, physical simulation, and differentiable rendering, we achieve 3D dynamic reconstruction and material estimation of fabrics.
Matic FuΔka (University of Ljubljana), Danijel SkoΔaj (University of Ljubljana)
CodeAnomaly DetectionImage
π― What it does: A logic anomaly detection framework named SALAD is proposed, which incorporates a composition branch that explicitly models the distribution of object compositions and achieves the discrimination of logical anomalies through synthetic anomaly training.
π― What it does: A transferable adversarial attack method called VeSCA is proposed, which utilizes the transferability of the SAM encoder to generate adversarial samples by constructing an Adversarial Simplicial Complex, thereby assessing the security of SAM in various downstream tasks.
SAM2Long: Enhancing SAM 2 for Long Video Segmentation with a Training-Free Memory Tree
Shuangrui Ding (Chinese University of Hong Kong), Jiaqi Wang (Shanghai AI Laboratory)
CodeObject TrackingSegmentationVideoBenchmark
π― What it does: This paper proposes SAM2Long, a training-independent improvement scheme that enhances the robustness and accuracy of SAM2 in long video segmentation through tree-shaped memory and uncertainty handling.
SAME: Learning Generic Language-Guided Visual Navigation with State-Adaptive Mixture of Experts
Gengze Zhou (University of Adelaide), Qi Wu (University of Adelaide)
CodeRobotic IntelligenceTransformerMixture of ExpertsVision Language ModelMultimodality
π― What it does: A general language-guided visual navigation model called SAME is proposed, which is based on state-adaptive expert mixture and can complete seven different granularity navigation tasks within a single framework.
SAMO: A Lightweight Sharpness-Aware Approach for Multi-Task Optimization with Joint Global-Local Perturbation
Hao Ban (University at Buffalo), Kaiyi Ji (University at Buffalo)
CodeOptimizationImage
π― What it does: A lightweight SAMO method is proposed, utilizing joint global-local perturbations to improve the optimization process of multi-task learning (MTL).
π― What it does: For medical image segmentation, the Segment Anything Model (SAM) is fine-tuned without prompts through multi-level self-supervised pre-training and low-rank adaptation (LoRA), and a hierarchical attention fusion module (HL-Attn) is introduced to enhance segmentation accuracy.
SC-Captioner: Improving Image Captioning with Self-Correction by Reinforcement Learning
Lin Zhang (Fudan University), Tao Chen (Fudan University)
CodeGenerationTransformerReinforcement LearningVision Language ModelImageText
π― What it does: This paper proposes SC-Captioner, a multi-round self-correction framework based on reinforcement learning, which enables large vision-language models to automatically eliminate false information and supplement missing details after generating image descriptions.
Scalable Dual Fingerprinting for Hierarchical Attribution of Text-to-Image Models
Jianwei Fei (University of Macau), Zhihua Xia (Jinan University)
CodeGenerationData SynthesisImageText
π― What it does: A scalable dual fingerprint embedding method has been developed for hierarchical attribution in text-to-image models, achieving traceability for both service providers and consumers.
π― What it does: This study investigates the issue of unified image generation models easily overlooking instructions when handling multiple sub-instructions, and proposes an adaptive attention scaling method to improve instruction following quality.
π― What it does: This paper explores the performance of visual self-supervised learning in multimodal tasks (VQA) by training visual self-supervised models (DINOv2, MAE) on the large-scale network image dataset MetaCLIP and expanding the model parameters from 1B to 7B, comparing it with CLIP.
π― What it does: The MiCo framework is proposed, utilizing large-scale multimodal alignment data (images, depth, normal maps, audio, video, text) for omni-modal pre-training, aiming to construct general and transferable cross-modal representations.
π― What it does: This study investigates the data scale rule for tumor segmentation and constructs the largest and most comprehensively annotated abdominal tumor dataset, AbdomenAtlas 2.0. It also verifies the role of synthetic data in improving performance both within and outside the distribution by synthesizing tumors to accelerate model training.
π― What it does: A dynamic guided contrastive pre-training data pruning method called SCAN is proposed, which can adaptively select and remove unimportant or redundant data samples during the pre-training process.
π― What it does: This paper constructs the SCIVID benchmark, covering five scientific video tasks including animal behavior classification, tissue tracking, weather and pressure prediction, and systematically evaluates the transfer effects of various video foundation models (ViFMs) on these tasks;
π― What it does: This paper proposes the SCORE framework, which implements open vocabulary-based remote sensing instance segmentation and enhances segmentation accuracy through multi-granularity scene context.
SDMatte: Grafting Diffusion Models for Interactive Matting
Longfei Huang (Shanghai University), Peng-Tao Jiang (vivo Mobile Communication Co)
CodeSegmentationGenerationDiffusion modelImage
π― What it does: This paper proposes an interactive image matting method called SDMatte based on diffusion models, which utilizes visual prompts (points, boxes, masks) to achieve precise Alpha Matte prediction.
Ivan SaboliΔ (University of Zagreb), SiniΕ‘a Ε egviΔ (University of Zagreb)
CodeOptimizationAdversarial AttackImage
π― What it does: A model-agnostic VIBE framework is proposed, treating clean labels as latent variables through variational inference, and recovering pseudo-labels during training using the EM algorithm and entropy-regularized optimal transport to resist backdoor attacks.
π― What it does: This paper proposes SeaS, a unified industrial anomaly image generation model that can simultaneously generate various anomalies, realistic normal products, and precise pixel-level anomaly masks.
π― What it does: A SEHDR framework is proposed, utilizing single-exposure multi-view LDR images to achieve HDR new view synthesis through 3D Gaussian splatting.
π― What it does: This paper proposes a zero-shot image denoising framework called Noise2VST, which utilizes self-supervised learning VST and a pre-trained Gaussian denoiser to achieve real-world noise removal.
π― What it does: A semi-supervised lifelong person re-identification method SPRED is proposed, which enhances the utilization of unlabeled data through self-reinforcement prototype evolution and dual knowledge collaboration.
π― What it does: A no-data quantization method for visual Transformers, SARDFQ, is proposed to address the issues of semantic distortion and semantic insufficiency in synthetic images.
π― What it does: This paper proposes an end-to-end supervised 2Dβ3D semantic occupancy prediction framework based on causal loss, addressing the error propagation caused by semantic confusion in the traditional Lift-Splat-Shoot method.
Semantic Discrepancy-aware Detector for Image Forgery Identification
Ziye Wang (Nanjing University of Science and Technology), Zhen Cui (Beijing Normal University)
CodeAnomaly DetectionTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImage
π― What it does: A semantic difference-aware image forgery detection method SDD is proposed, which aligns the semantic concepts of the CLIP visual-language model with forgery features and enhances detection performance through reconstruction learning.
Sequential keypoint density estimator: an overlooked baseline of skeleton-based video anomaly detection
Anja DeliΔ (University of Zagreb), SiniΕ‘a Ε egviΔ (University of Zagreb)
CodeAnomaly DetectionTransformerVideoSequential
π― What it does: This paper proposes a skeleton sequence anomaly detection method called SeeKer, which identifies abnormal behaviors by predicting the probability distribution of skeletal key points.
π― What it does: A point cloud refinement method SPCNet based on Hilbert curve serialization is designed, utilizing serialized initial segmentation, semantic similarity matching, and cross-attention to achieve superpoint generation and iterative optimization.
π― What it does: A semantic-guided LiDAR diffusion model SG-LDM is proposed to directly convert semantic segmentation maps into high-quality LiDAR point clouds.
ShadowHack: Hacking Shadows via Luminance-Color Divide and Conquer
Jin Hu (Tianjin University), Xiaojie Guo (Tianjin University)
CodeRestorationTransformerImage
π― What it does: A method is proposed to decompose shadow images into brightness and color components, and to perform brightness recovery and color reconstruction using LRNet and CRNet, respectively, to achieve shadow removal.
ShortV: Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers
Qianhao Yuan (Institute of Software, Chinese Academy of Sciences), Le Sun (Institute of Software, Chinese Academy of Sciences)
CodeComputational EfficiencyTransformerLarge Language ModelVision Language ModelMultimodality
π― What it does: This paper proposes a method called ShortV, which is untrained and freezes visual tokens to enhance the reasoning efficiency of multimodal large language models.
π― What it does: A medical image segmentation network called Sim-MPNet based on similarity memory priors is proposed, along with two core modules: Dynamic Memory Weights-Loss Attention (DMW-LA) and Double-Similarity Global Internal Enhancement Module (DS-GIM).
SmolDocling: An ultra-compact vision-language model for end-to-end multi-modal document conversion
Ahmed Nassar (IBM Research), Peter W. J. Staar (IBM Research)
CodeRecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityTabular
π― What it does: We propose SmolDocling, a 256M parameter visual-language model that achieves end-to-end multimodal document conversion, outputting a unified DocTags markup format.
SMoLoRA: Exploring and Defying Dual Catastrophic Forgetting in Continual Visual Instruction Tuning
Ziqi Wang (Hefei University of Technology), Meng Wang (Hefei University of Technology)
CodeClassificationRecognitionData-Centric LearningTransformerLarge Language ModelMixture of ExpertsVision Language ModelImageTextMultimodalityBenchmark
π― What it does: This paper addresses the problem of Continuous Visual Instruction Tuning (CVIT) for multimodal large language models and proposes the SMoLoRA framework, which utilizes separable routing (between the visual understanding module and the instruction following module) and adaptive fusion to tackle dual catastrophic forgetting, and constructs a CVIT benchmark covering multiple tasks, multiple instructions, and unseen tasks.
π― What it does: A new transferable adversarial attack method called SMP-Attack is proposed, which integrates semantic-aware multi-granularity patch replacement and multi-stage optimization to enhance the attack success rate of feature importance attacks in black-box scenarios.
π― What it does: Proposes SMSTrackerβa three-path multimodal tracking framework that integrates RGB with thermal/depth/event multimodal information, including modules such as Score Mask Fusion, Sigma Interaction, and Drop Key Fine-tuning.
CodeExplainability and InterpretabilityConvolutional Neural NetworkTransformerImage
π― What it does: The SLOC method is proposed, which softens the integrity constraints on sub-regions of the attribution map and achieves more trustworthy interpretability maps by minimizing the integrity gap.
π― What it does: This paper proposes a dual-stream point transformer SP2T based on sparse proxy attention, aimed at achieving a global receptive field in 3D point clouds while maintaining detail extraction.
π― What it does: This paper proposes an interactive medical image segmentation framework named SPA, which can efficiently generate multiple representative segmentation results during the inference phase based on user preferences in different clinical contexts and converge quickly.
π― What it does: We propose SparseFlex, a differentiable isosurface representation with a sparse structure, and build a VAE generation pipeline based on this representation, achieving high-resolution arbitrary topology 3D shape modeling and single-image 3D generation.
SparseMM: Head Sparsity Emerges from Visual Concept Responses in MLLMs
Jiahui Wang (Tsinghua University), Jiwen Lu (Tsinghua University)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelTextMultimodality
π― What it does: By analyzing the attention mechanism of multimodal large language models (MLLMs), it reveals that visual-related attention heads are extremely sparse (<5%), and based on this, SparseMM is proposedβa head-level KV-Cache dynamic allocation strategy to achieve efficient acceleration.
Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation
Nairouz Mrabah (ETS Montreal), Eric Granger (ETS Montreal)
CodeDomain AdaptationOptimizationSupervised Fine-TuningVision Language ModelImageMultimodality
π― What it does: This paper proposes a sparse optimization (SO) framework for efficiently fine-tuning visual-language models (VLMs) in scenarios with very few labeled samples.
Spatial Preference Rewarding for MLLMs Spatial Understanding
Han Qiu (Nanyang Technological University), Shijian Lu (Nanyang Technological University)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningImageMultimodality
π― What it does: By constructing random image regions and utilizing preference optimization to enhance the capability of multimodal large language models (MLLM) for fine-grained spatial understanding.
π― What it does: A framework for image forgery detection based on spatio-temporal forged traces (STFT) has been developed, utilizing the temporal distribution of latent diffusion models to capture forgery features;
SPD: Shallow Backdoor Protecting Deep Backdoor Against Backdoor Detection
Shunjie Yuan (Xidian University), Robert H. Deng (Singapore Management University)
CodeAdversarial AttackAuto EncoderImage
π― What it does: A white-box backdoor attack method called SPD is designed, combining shallow pixel domain triggers with deep frequency domain amplitude triggers to achieve the concealment and defense detection of backdoors.
π― What it does: A spike neural network based on Transformer, called SpiLiFormer, is proposed, which improves spike attention using a brain-inspired lateral inhibition mechanism to enhance image classification performance.
π― What it does: This paper proposes a self-supervised two-state RGB image recovery method called SPLART, which simultaneously reconstructs and infers the motion model of movable parts using 3D Gaussian Splatting, achieving real-time and realistic viewpoint synthesis.
Splat-based 3D Scene Reconstruction with Extreme Motion-blur
Hyeonjoong Jang (Korea Advanced Institute of Science and Technology), Min H. Kim (Korea Advanced Institute of Science and Technology)
CodeRestorationPose EstimationDepth EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingOptical FlowImagePoint Cloud
π― What it does: A robust 3D scene reconstruction method based on RGB-D input is proposed, capable of simultaneously achieving camera pose estimation, point cloud refinement, and image deblurring under extreme motion blur.
SRefiner: Soft-Braid Attention for Multi-Agent Trajectory Refinement
Liwen Xiao (Huazhong University of Science and Technology), Wei Li (Nanyang Technological University)
CodeAutonomous DrivingTime Series
π― What it does: A multi-agent trajectory refinement framework SRefiner based on soft junction topology is proposed, which captures the spatiotemporal topological relationship between trajectories and lanes through soft junctions, and continuously updates the topological information in multiple iterations to improve prediction accuracy.
π― What it does: Improving the limitations of traditional vector quantization during the fine-tuning phase, a learnable symbol-split vector quantization (SSVQ) is proposed, allowing each quantization weight to independently follow gradient updates;
π― What it does: A Stable Score Distillation (SSD) method is proposed for text-guided image and 3D scene editing, which achieves more realistic editing effects while maintaining the original content structure.
STaR: Seamless Spatial-Temporal Aware Motion Retargeting with Penetration and Consistency Constraints
Xiaohang Yang (Queen Mary University of London), Shanxin Yuan (Queen Mary University of London)
CodeTransformerPoint CloudMesh
π― What it does: STaR proposes an end-to-end spatiotemporal motion remapping framework that can generate complete, geometrically reasonable, and temporally coherent motion for target characters from the motion sequences of source characters.
Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images
Qi Xun Yeo (National University of Singapore), Gim Hee Lee (National University of Singapore)
CodeObject DetectionSegmentationGenerationRetrievalGraph Neural NetworkSimultaneous Localization and MappingImagePoint Cloud
π― What it does: A framework for generating 3D semantic scene graphs based on multi-view RGB images is proposed, which mainly includes modules for initializing node features using segmentation masks, improving edge features based on residual neighborhood graph convolutional networks, and confidence re-labeling using statistical priors.
π― What it does: A spatiotemporal decoupling framework for high dynamic scene reconstruction, STD-GS, is proposed based on event camera assistance. It first decomposes dynamic scenes into background and dynamic objects, then fuses spatial color and temporal deformation on 4D Gaussian points, ultimately achieving temporally continuous rendering of high dynamic scenes.
π― What it does: This paper proposes STDDNet, a video polyp segmentation network that combines spatially aligned temporal modeling with discriminative dynamic representation learning.
π― What it does: This paper proposes an unsupervised generative framework that utilizes stochastic interpolation and dual diffusion implicit bridges, capable of mapping a single artwork to different historical style distributions without paired data, thereby revealing the evolution of artistic styles over time.
π― What it does: An end-to-end real-time streaming diffusion pipeline, StreamDiffusion, is proposed to achieve high-throughput interactive image generation.
π― What it does: This paper studies an untrained sketch generation framework called Stroke2Sketch, which can accurately transfer the style of content images using the stroke attributes of reference sketches while maintaining the integrity of the semantic structure.
π― What it does: DepthForge fine-tunes domain generalization semantic segmentation by integrating a pre-trained depth VFM with a visual VFM, utilizing depth-aware learnable tokens.
π― What it does: This paper proposes a structure-aware 3D medical image self-supervised learning framework SDC2, aimed at learning the semantic differences and consistencies between different anatomical structures.
π― What it does: A shadow removal method for portraits based on diffusion models is proposed, treating shadow removal as a structure-guided filling task.
π― What it does: A style-embedded scene text image super-resolution network, StyleSRN, is proposed to simultaneously restore text structure and style.
π― What it does: A Stylized-Face dataset (4.6M images, 62k IDs, 80 styles) was constructed and three sets of evaluation benchmarks were proposed for studying generative style face recognition.