IEEE/CVF International Conference on Computer Vision · 833 papers
Bring Your Rear Cameras for Egocentric 3D Human Pose Estimation
Hiroyasu Akada (Max Planck Institute for Informatics), Christian Theobalt (Max Planck Institute for Informatics)
CodePose EstimationTransformerImage
🎯 What it does: Introducing a rear camera in an egocentric environment for full-body 3D pose estimation, and proposing a Transformer-based heatmap refinement module to improve estimation accuracy.
BUFFER-X: Towards Zero-Shot Point Cloud Registration in Diverse Scenes
Minkyun Seo (Seoul National University), Jaesik Park (Seoul National University)
CodePose EstimationOptimizationSimultaneous Localization and MappingPoint CloudBenchmark
🎯 What it does: A completely zero-shot point cloud registration pipeline called BUFFER-X is proposed, which can achieve global registration in diverse scenarios without any manual parameter tuning or training data transfer.
C2MIL: Synchronizing Semantic and Topological Causalities in Multiple Instance Learning for Robust and Interpretable Survival Analysis
Min Cen (University of Science and Technology of China), Liansheng Wang (Xiamen University)
CodeClassificationExplainability and InterpretabilityGraph Neural NetworkTransformerContrastive LearningBiomedical Data
🎯 What it does: A dual causal graph-based multi-instance learning framework, C MIL, is proposed to eliminate semantic bias in WSI and identify causal subgraphs to enhance survival analysis.
CA2C: A Prior-Knowledge-Free Approach for Robust Label Noise Learning via Asymmetric Co-learning and Co-training
Mengmeng Sheng (Nanjing University of Science and Technology), Yazhou Yao (Beijing Institute of Technology)
CodeClassificationData-Centric LearningImage
🎯 What it does: This paper proposes a robust label noise learning framework CA2C without prior knowledge, utilizing asynchronous co-learning and co-training strategies of two models to achieve adaptive suppression of noise.
🎯 What it does: A framework for contrast-invariant self-supervised landmark detection, CABLD, is proposed, which requires only a single template annotation to perform detection in unannotated 3D brain MR scans.
CAD-Assistant: Tool-Augmented VLLMs as Generic CAD Task Solvers
Dimitrios Mallis (University of Luxembourg), Djamila Aouada (University of Luxembourg)
CodeLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Proposes CAD-Assistant, a general CAD agent based on Vision-Large-Language-Model, which implements multimodal question answering and editing using the FreeCAD API;
Can Generative Geospatial Diffusion Models Excel as Discriminative Geospatial Foundation Models?
Yuru Jia (KU Leuven), Andrea Nascetti (KTH)
CodeClassificationSegmentationMixture of ExpertsDiffusion modelImageBenchmark
🎯 What it does: Transform the generative diffusion model into a self-supervised remote sensing foundation model, and perform pre-training and fine-tuning on various discriminative tasks.
🎯 What it does: A scene-level variational autoencoder Can3Tok based on 3D Gaussian splatting is designed to achieve low-dimensional latent encoding and reconstruction of large-scale scenes.
CaptionSmiths: Flexibly Controlling Language Pattern in Image Captioning
Kuniaki Saito (OMRON SINIC X), Yoshitaka Ushiku
CodeGenerationTransformerLarge Language ModelVision Language ModelImageText
🎯 What it does: A single image captioning model called CaptionSmiths is proposed, which can continuously control the length, descriptiveness (information density), and vocabulary uniqueness of captions by inserting interpolable conditional vectors into the input of the language model.
CAPTURE: Evaluating Spatial Reasoning in Vision Language Models via Occluded Object Counting
Atin Pothiraj (University of North Carolina Chapel Hill), Mohit Bansal (University of North Carolina Chapel Hill)
CodeObject DetectionTransformerVision Language ModelDiffusion modelImageMultimodalityBenchmark
🎯 What it does: This paper proposes the CAPTURE benchmark to evaluate the ability of visual language models to count modal patterns in occluded scenes.
CARL: Causality-guided Architecture Representation Learning for an Interpretable Performance Predictor
Han Ji (Sichuan University), Yanan Sun (Sichuan University)
CodeExplainability and InterpretabilityRepresentation LearningNeural Architecture SearchGraph Neural NetworkImage
🎯 What it does: This paper proposes a causal intervention-based architecture representation learning method called CARL, which enhances the generalization and interpretability of performance predictors in neural architecture search by splitting key and redundant features in the latent space and generating cross-intervention samples.
🎯 What it does: This paper proposes CasP—a semi-dense feature matching pipeline based on cascade correspondence priors, aimed at improving matching speed and robustness.
CAT: A Unified Click-and-Track Framework for Realistic Tracking
Yongsheng Yuan (Dalian University of Technology), Huchuan Lu (Dalian University of Technology)
CodeObject TrackingTransformerMixture of ExpertsVideo
🎯 What it does: A unified click-tracking framework CAT is proposed, which utilizes a single point click to complete target initialization and achieve continuous tracking.
Category-Specific Selective Feature Enhancement for Long-Tailed Multi-Label Image Classification
Ruiqi Du (Xidian University), Jingjing Ma (Xidian University)
CodeClassificationRecognitionTransformerImage
🎯 What it does: This paper proposes a framework based on Category-Specific Selective Feature Enhancement (CSSFE) for the long-tail multi-label image classification task, aiming to improve the representation and recognition capabilities of rare categories.
Causality-guided Prompt Learning for Vision-language Models via Visual Granulation
Mengyu Gao (Chinese Academy of Sciences), Qiulei Dong (Chinese Academy of Sciences)
CodeClassificationRecognitionDomain AdaptationTransformerPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes CaPL, a causal-guided text prompt learning method that enhances CLIP's performance on fine-grained tasks by utilizing visual fine-grained decomposition and visual granulation techniques.
CE-FAM: Concept-Based Explanation via Fusion of Activation Maps
Michihiro Kuroki (University of Tokyo), Toshihiko Yamasaki (University of Tokyo)
CodeExplainability and InterpretabilityVision Language ModelImage
🎯 What it does: A concept-level explainable method CE-FAM is proposed, which can simultaneously identify the concepts learned by image classification models, locate their corresponding areas, and quantify their contributions to predictions.
🎯 What it does: This paper proposes a provably optimal rotation averaging method under anisotropic uncertainty, which explicitly incorporates the confidence information of each relative rotation into the optimization objective.
🎯 What it does: This paper studies the robustness of monocular 3D detection in the absence of seen camera height (ego height) and proposes the CHARM3R model based on average depth estimation, significantly improving detection performance at different heights.
🎯 What it does: By using multi-teacher adversarial distillation, the robustness and accuracy of the teacher model are transferred to a lightweight student model, proposing the CIARD framework.
Client2Vec: Improving Federated Learning by Distribution Shifts Aware Client Indexing
Yongxin Guo (Westlake University), Tao Lin (Westlake University)
CodeFederated LearningSafty and PrivacyContrastive LearningImageText
🎯 What it does: Proposes the Client2Vec mechanism, which generates an index vector containing distribution shift information for each client before federated learning training, and uses this index to improve client sampling, model aggregation, and local training;
CLOT: Closed Loop Optimal Transport for Unsupervised Action Segmentation
Elena Bueno-Benito (Institut de Robótica y Informática Industrial), Mariella Dimiccoli (Institut de Robótica y Informática Industrial)
CodeSegmentationOptimizationVideo
🎯 What it does: This paper studies a new framework for unsupervised action segmentation called CLOT, which optimizes frame and segment representations through closed-loop optimal transport.
🎯 What it does: The CMAD framework is proposed, which achieves unified representation in the task of missing multimodal emotion analysis through teacher-student knowledge distillation, and designs two modules: CAFD for sample-level feature and relevance alignment, and MAR for modality-aware regularization.
🎯 What it does: This paper proposes CO‑2Net, a physics-informed spatiotemporal model designed to reconstruct global surface CO2 concentrations from sparse observations and meteorological auxiliary variables, avoiding reliance on large-scale prior data.
CODE-CL: Conceptor-Based Gradient Projection for Deep Continual Learning
Marco P. E. Apolinario (Purdue University), Kaushik Roy (Purdue University)
CodeClassificationOptimizationImage
🎯 What it does: A gradient projection method based on Conceptor, CODE-CL, is proposed to simultaneously suppress catastrophic forgetting and enhance forward knowledge transfer in continual learning.
CodeAutonomous DrivingTransformerLarge Language ModelVision Language ModelTextBenchmark
🎯 What it does: This paper proposes CoLMDriver, a complete LLM-driven collaborative driving system that combines multi-turn language negotiation and intent-guided waypoint generation, and introduces the InterDrive interactive driving benchmark.
🎯 What it does: This paper proposes the cmKAN framework for mapping the colors of source images to target color spaces, supporting three scenarios: supervised, unsupervised, and paired optimization.
Colors See Colors Ignore: Clothes Changing ReID with Color Disentanglement
Priyank Pathak (University of Central Florida), Yogesh S. Rawat (University of Central Florida)
CodeRecognitionRetrievalTransformerImageVideo
🎯 What it does: Proposes a clothing change re-identification method called CSCI that only uses RGB without external annotations or models, utilizing color information to decouple identity features;
🎯 What it does: A semi-dense image matcher called CoMatch is designed, which combines a dynamic visibility-aware Transformer and a bidirectional sub-pixel refinement module to achieve high-precision and high-efficiency image matching.
COME: Dual Structure-Semantic Learning with Collaborative MoE for Universal Lesion Detection Across Heterogeneous Ultrasound Datasets
Lingyu Chen (Nanjing University of Aeronautics and Astronautics), Fang Chen (Nanjing University of Aeronautics and Astronautics)
CodeObject DetectionMixture of ExpertsImageMultimodalityBiomedical DataUltrasound
🎯 What it does: This paper studies a cross-homogeneous ultrasound imaging multi-dataset general lesion detection framework based on dual-structure semantic learning and Collaborative Mixture of Experts (COME).
🎯 What it does: A communication-efficient multi-vehicle collaborative semantic segmentation framework GSCOOP based on sparse 3D Gaussian sharing has been developed, which can generate discrete 3D Gaussian representations from multi-view images and achieve low-bandwidth communication through selection and compression.
CompCap: Improving Multimodal Large Language Models with Composite Captions
Xiaohui Chen (Meta), Baosheng He (Meta)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningImageTextMultimodality
🎯 What it does: Proposes the CompCap framework, which utilizes LLM and automated tools to generate six types of synthetic images and their high-quality, detailed titles;
🎯 What it does: A first-order diffusion model CODiff is proposed, which uses a compressed sensing visual embedder CaVE to extract JPEG compression priors, and based on this, performs artifact removal and reconstruction on low-quality images.
🎯 What it does: This paper proposes the Object Conditional Diffusion Transformer (OC-DiT), a zero-shot instance segmentation method based on diffusion models that can generate instance masks without the need for training on target objects.
Justin Kay (Massachusetts Institute of Technology), Sara Beery (Massachusetts Institute of Technology)
CodeClassificationBenchmark
🎯 What it does: A consensus-driven active model selection method called CODA is proposed, which quickly identifies the best pre-trained model using a small number of labels.
🎯 What it does: A model for generating super-resolution images in a single step without distillation (CTMSR) is proposed, which learns a deterministic mapping from low-resolution images with noise to high-resolution images through consistency training, and further enhances the naturalness of generated images using Distribution Trajectory Matching (DTM).
🎯 What it does: This paper proposes a time-of-flight (ToF) depth denoising network based on cross-frame graph structure fusion—GIGA-ToF, which enhances temporal consistency while preserving spatial details.
Context Guided Transformer Entropy Modeling for Video Compression
Junlong Tong (Shanghai Jiao Tong University), Xiaoyu Shen (Ningbo Key Laboratory of Spatial Intelligence and Digital Derivative, Institute of Digital Twin, EIT)
CodeCompressionTransformerVideo
🎯 What it does: Proposes the Context Guided Transformer (CGT) conditional entropy model to improve the accuracy and efficiency of entropy modeling in video compression.
🎯 What it does: A 3D outdoor scene generation framework based on scene graph control is proposed, which includes an interactive system, BEV embedded graph, and conditional diffusion model, enabling the automatic generation of user-editable scene graphs into complete 3D urban landscapes.
🎯 What it does: A framework called FedCoPL is proposed for classification in unlabeled federated learning using CLIP, combining collaborative pseudo-label generation and local prompt aggregation.
🎯 What it does: An end-to-end cooperative 3D multi-object tracking framework called CoopTrack is proposed, achieving a complete pipeline of instance-level collaborative learning and post-decoding fusion.
🎯 What it does: This paper proposes a correspondence-free spherical point pattern registration algorithm that can quickly estimate the rotational relationship between two sets of spherical point clouds.
CoST: Efficient Collaborative Perception From Unified Spatiotemporal Perspective
Zongheng Tang (Beihang University), Si Liu (Beihang University)
CodeAutonomous DrivingComputational EfficiencySimultaneous Localization and MappingPoint Cloud
🎯 What it does: The CoST framework is proposed to achieve unified collaborative perception for multiple vehicles at multiple time points, efficiently fusing information in spatial and temporal dimensions through the STT and USTF modules.
🎯 What it does: A new open set recognition method called COSTARR is proposed, which determines whether a sample belongs to a known category by combining pre-decayed features (original deep features) and post-decayed features (Hadamard product) along with normalization of the values.
COVTrack: Continuous Open-Vocabulary Tracking via Adaptive Multi-Cue Fusion
Zekun Qian (Tianjin University), Wei Feng (City University of Hong Kong)
CodeObject TrackingKnowledge DistillationVideo
🎯 What it does: This paper proposes the C-TAO dataset with continuous annotations and builds the COVTrack framework based on it for open vocabulary multi-object tracking.
Creation-MMBench: Assessing Context-Aware Creative Intelligence in MLLMs
Xinyu Fang (Zhejiang University), Dahua Lin (Zhejiang University)
CodeTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Created Creation-MMBench to evaluate the performance of multimodal large language models in visual creative tasks, including 765 test cases, 51 fine-grained tasks, and providing a text version Creation-MMBench-TO.
🎯 What it does: A cross-architecture knowledge distillation method called RSD is proposed, which extracts architecture-independent knowledge through redundancy suppression and transfers it to the student network.
Cross-modal Ship Re-Identification via Optical and SAR Imagery: A Novel Dataset and Method
Han Wang (Technology and Engineering Center for Space Utilization, Chinese Academy of Sciences), Zhuang Zhou (Technology and Engineering Center for Space Utilization, Chinese Academy of Sciences)
🎯 What it does: A cross-modal ship re-identification dataset HOSS ReID has been established, and a cross-modal re-identification method based on Vision Transformer, TransOSS, has been proposed.
CT-ScanGaze: A Dataset and Baselines for 3D Volumetric Scanpath Modeling
Trong Thang Pham (University of Arkansas), Ngan Le (University of Liverpool)
CodeExplainability and InterpretabilityTransformerBiomedical DataComputed Tomography
🎯 What it does: This paper presents the first publicly available CT eye movement dataset, CT-ScanGaze, and proposes a 3D eye movement path prediction model, CT-Searcher, which further opens new directions for explainable AI in medical imaging.
🎯 What it does: This paper proposes a Custom Domain Adapter (CDA) framework for domain generalization, utilizing lightweight ViT and CNN adapters to focus on learning the data features of their respective domains, and dynamically merging predictions through a domain router;
🎯 What it does: This paper designs a two-stage cross-view fusion network called CVFusion, which fuses 4D radar point clouds with camera images for 3D object detection.
🎯 What it does: The GDCC (Generation-Detection Cycle-Consistent) framework is proposed, which simultaneously optimizes the layout-to-image (L2I) and object detection (OD) tasks in an end-to-end training process.
🎯 What it does: This study investigates the differences in second-order temporal features (acceleration) between AI-generated videos and real videos, proposing a training-free detection method called D3 based on second-order central differences.
🎯 What it does: A method based on discrete distribution difference quantization error (D QE 3) is proposed to detect images generated by visual autoregressive models.
🎯 What it does: A deep angle A* (DAA*) method is proposed to improve image-based path planning by learning the Path Angle Freedom (PAF), making the predicted paths closer to expert demonstrations and achieving better smoothness.
🎯 What it does: DACoN is designed to colorize animation line art using DINOv2 and U-Net with multiple reference images, taking into account both key frames and continuous frames.
DALIP: Distribution Alignment-based Language-Image Pre-Training for Domain-Specific Data
Junjie Wu (Tianjin University), Sen Xu (Dalian University of Technology)
CodeClassificationDomain AdaptationRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkAgriculture Related
🎯 What it does: This paper proposes a distribution alignment-based CLIP pre-training method (DALIP), which aligns the model using the first and second-order statistics of image-text feature distributions, and constructs the PlantMix-13M plant-specific dataset to further enhance the model's performance in the fine-grained biological domain while maintaining compatibility with the general domain.
DASH: 4D Hash Encoding with Self-Supervised Decomposition for Real-Time Dynamic Scene Rendering
Jie Chen (University of Science and Technology of China), Xiaoyan Sun (Institute of Artificial Intelligence, Hefei Comprehensive National Science Center)
🎯 What it does: A real-time dynamic scene rendering framework named DASH is proposed, utilizing 4D hash encoding and self-supervised dynamic-static decomposition to achieve high-quality, real-time rendering;
🎯 What it does: This paper proposes a dataset distillation method based on Wasserstein barycenter (WMDD), which optimizes synthetic samples in the feature space of a pre-trained model to align their distribution with the Wasserstein barycenter of the real data.
🎯 What it does: This paper proposes DC-AE 1.5, a deep compression autoencoder with a structured latent space and enhanced diffusion training, aimed at accelerating the convergence of latent diffusion models with a large number of channels and improving the quality of high-resolution image generation.
DC-TTA: Divide-and-Conquer Framework for Test-Time Adaptation of Interactive Segmentation
Jihun Kim (KAIST), Kuk-Jin Yoon (KAIST)
CodeSegmentationDomain AdaptationImage
🎯 What it does: A divide-and-conquer testing-time adaptation framework DC-TTA is proposed to improve the performance of the interactive segmentation model SAM.
🎯 What it does: A method for imperceptible adversarial perturbation of images in the DCT domain is designed, utilizing the JPEG quantization process to protect against malicious editing of diffusion models.
DDB: Diffusion Driven Balancing to Address Spurious Correlations
Aryan Yazdan Parast (University of Melbourne), Naveed Akhtar (University of Melbourne)
CodeGenerationData SynthesisDiffusion modelImage
🎯 What it does: Automatically generate minority class samples using text inversion, language segmentation, and local inpainting of Stable Diffusion to rebalance the training set and reduce the model's reliance on spurious correlations.
Debiased Curriculum Adaptation for Safe Transfer Learning in Chest X-ray Classification
Mingyang Liu (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
CodeClassificationDomain AdaptationImageBiomedical Data
🎯 What it does: A Debiased Curriculum Adaptation framework is proposed to achieve safe unsupervised domain adaptation in chest X-ray image classification tasks.
Decoding Correlation-Induced Misalignment in the Stable Diffusion Workflow for Text-to-Image Generation
Yunze Tong (Zhejiang University), Kun Kuang (Zhejiang University)
CodeGenerationData SynthesisTransformerVision Language ModelDiffusion modelImageText
🎯 What it does: To address the 'object missing' problem in Stable Diffusion when handling related words in text, a de-correlation fine-tuning method is proposed on the self-attention layer of the text encoder to improve the alignment between text and images without introducing external knowledge.
🎯 What it does: A decoupled multi-predictor optimization (DMPO) method based on early exit is proposed to achieve efficient fine-tuning of inference for large-scale pre-trained models.
🎯 What it does: A deep adaptive unfolding network (DAPNet) is proposed, which achieves multi-modal image fusion (full-resolution spectral image reconstruction) through spatial morphology stripping and spectral filtering.
🎯 What it does: This paper proposes a depth enhancement method for dToF sensors based on deep completion, called DepthOR, which includes a noise-robust training strategy and a two-stage network that incorporates monocular depth estimation.
Derm1M: A Million-scale Vision-Language Dataset Aligned with Clinical Ontology Knowledge for Dermatology
Siyuan Yan (Monash University), Zongyuan Ge (Monash University)
CodeClassificationRecognitionRetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityAudio
🎯 What it does: This paper constructs Derm1M, the first large-scale audiovisual language dataset containing 1,029,761 pairs of skin images and text, 390 types of skin diseases, and 130 clinical concepts, and pre-trains the DermLIP visual-language model based on this dataset.
Describe, Adapt and Combine: Empowering CLIP Encoders for Open-set 3D Object Retrieval
Zhichuan Wang (Huazhong Agricultural University), Xiang Bai (Huazhong University of Science and Technology)
CodeRetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningImageMultimodalityPoint Cloud
🎯 What it does: This paper proposes the DAC framework, which utilizes multi-view images in combination with CLIP and a multimodal large language model to achieve open-set 3D object retrieval.
🎯 What it does: The DeSPITE model is proposed, which maps LiDAR point clouds, skeletons, IMU, and text into a shared embedding space to achieve cross-modal matching, retrieval, and pre-training for human action recognition.
🎯 What it does: This paper proposes DetAny3D, a foundational model that enables 3D detection on any monocular image through prompts such as boxes, points, and text.
Devil is in the Uniformity: Exploring Diverse Learners within Transformer for Image Restoration
Shihao Zhou (Nankai University), Jufeng Yang (Nankai University)
CodeRestorationTransformerImage
🎯 What it does: For the image restoration task, a Transformer-based model called HINT is proposed, which achieves image denoising, dehazing, deraining, desnowing, and low-light enhancement by improving the multi-head attention mechanism.
DictAS: A Framework for Class-Generalizable Few-Shot Anomaly Segmentation via Dictionary Lookup
Zhen Qu (Institute of Automation Chinese Academy of Sciences), Guiguang Ding (Tsinghua University)
CodeSegmentationAnomaly DetectionTransformerVision Language ModelContrastive LearningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes the DictAS framework, which rephrases the few-shot anomaly segmentation task as a dictionary lookup problem, enabling the detection of anomalies in unseen categories without retraining on the target data.
🎯 What it does: A fully differentiable image-to-point cloud (I2P) registration framework called Diff2I2P is proposed, which utilizes controlled side score distillation (CSD) from diffusion models and a differentiable registration solver BPnP during training to learn cross-modal features and correspondences.
DiffSim: Taming Diffusion Models for Evaluating Visual Similarity
Yiren Song (Show Lab National University of Singapore), Mike Zheng Shou (Show Lab National University of Singapore)
CodeRetrievalDiffusion modelImage
🎯 What it does: This paper proposes the DiffSim method, which utilizes the attention layers of the pre-trained diffusion model U-Net to extract features and calculates image similarity through the Aligned Attention Score.
🎯 What it does: This paper proposes an image-guided synthesis technique based on diffusion models, constructing a data interpolation curve from synthetic to real data, and designing an adaptive curriculum learning strategy based on this curve to enhance model robustness against low-quality or scarce data.
🎯 What it does: Proposes a training method for an unsupervised Selective Head Attention mechanism and improves the CLIP model for open vocabulary semantic segmentation;
🎯 What it does: This work proposes a variational codebook learning method based on the Dirichlet distribution, which continuousizes the codebook representation of traditional discrete VQ-VAE to achieve temporal consistency recovery of video faces.
DisCo: Towards Distinct and Coherent Visual Encapsulation in Video MLLMs
Jiahe Zhao (University of Chinese Academy of Sciences), Hengshuang Zhao (University of Hong Kong)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes DisCo, a method for achieving visual encapsulation in video multimodal large language models, capable of generating semantically clear and temporally consistent visual tokens.
Discontinuity-aware Normal Integration for Generic Central Camera Models
Francesco Milano (ETH Zurich), Robert Thiel (Meta)
CodeDepth EstimationOptimizationImageBenchmark
🎯 What it does: An explicit discrete point fusion method based on the local plane assumption and light ray direction is proposed to recover 3D depth from surface normal maps;
🎯 What it does: A CT reconstruction framework based on Discrete Gaussian Representation (DGR) is proposed, which can directly generate three-dimensional voxel volumes in an end-to-end manner.
🎯 What it does: This paper proposes a dual-stream structure called DISC, which learns instance and scene context separately through instance queries and scene queries in the BEV space, thereby achieving more refined 3D semantic scene completion.
🎯 What it does: A proactive defense method called PaRaMS is proposed to prevent other models from merging at the parameter level without affecting the original functionality of the model.
🎯 What it does: A dynamic iterative threshold shrinkage network (DISTA-Net) has been designed and implemented for separating densely arranged infrared small targets, achieving sub-pixel level demixing and radiance intensity estimation, and an open-source ecosystem has been released.
🎯 What it does: An end-to-end multimodal autonomous driving framework called DistillDrive is constructed, which enhances planning and decision-making performance through teacher model knowledge distillation, reinforcement learning, and generative models.
🎯 What it does: An Ensemble Parallel Direction (EPD) solver is proposed to reduce truncation errors in ODE sampling of diffusion models through parallel gradient evaluations, along with a plugin version called EPD-Plugin.
DisTime: Distribution-based Time Representation for Video Large Language Models
Yingsen Zeng (Meituan Inc), Yang Liu (Peking University)
CodeGenerationRetrievalTransformerLarge Language ModelVideo
🎯 What it does: Proposes the DisTime framework, which achieves continuous time representation using a single time marker, and implements event boundary localization through a distributed time decoder and encoder.
🎯 What it does: A post-training quantization framework DMQ is proposed, which utilizes learned equivalent scaling and per-channel binary scaling to address the outlier problem in the quantization of diffusion models.
🎯 What it does: Using a denoising diffusion model to predict five intrinsic properties of objects (albedo, metallic, roughness, normal, depth) from a single indoor RGB image.
DocThinker: Explainable Multimodal Large Language Models with Rule-based Reinforcement Learning for Document Understanding
Wenwen Yu (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
CodeExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningVision Language ModelTextMultimodalityChain-of-Thought
🎯 What it does: A rule-based reinforcement learning framework called DocThinker is proposed, which can dynamically improve the reasoning process of multimodal large language models (MLLMs) during inference and output interpretable intermediate steps (inference trajectories, restated questions, areas of focus, final answers).
DOGR: Towards Versatile Visual Document Grounding and Referring
Yinan Zhou (Xi'an Jiaotong University), Li Zhu (Xi'an Jiaotong University)
CodeRecognitionObject DetectionRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: The DOGR-Engine is proposed to generate high-quality document parsing and instruction tuning data, based on which DOGR-Bench and the DOGR model are constructed, significantly enhancing document localization, citation, and dialogue capabilities.
🎯 What it does: A domain-generalizable portrait style transfer framework is proposed, utilizing semantic correspondence and a dual conditional diffusion model to achieve high-quality, semantically aligned style transfer.
Domain-aware Category-level Geometry Learning Segmentation for 3D Point Clouds
Pei He (Xidian University), Wenping Ma (Xidian University)
CodeSegmentationDomain AdaptationPoint Cloud
🎯 What it does: A category-level geometric learning framework is proposed to enhance the domain generalization ability of point cloud semantic segmentation through category geometric embedding and geometric consistency learning.