π― What it does: Proposes CAV-MAE Sync, which improves audio-visual self-supervised learning through fine-grained audio-visual alignment, separation of contrastive and reconstruction objectives, and the introduction of global and register tokens.
CCIN: Compositional Conflict Identification and Neutralization for Composed Image Retrieval
Likai Tian (Wuhan University), Xuelong Li (Northwestern Polytechnical University)
CodeRetrievalTransformerLarge Language ModelContrastive LearningImage
π― What it does: This paper proposes a method to improve image retrieval accuracy by first identifying attribute conflicts between the reference image and modification instructions, and then neutralizing these conflicts through a dual-instruction mechanism.
π― What it does: This paper proposes a copyright data identification framework called CDI based on dataset inference, which is used to determine whether a Diffusion model has utilized a specific copyrighted dataset.
π― What it does: This paper proposes a new semi-supervised learning framework called CGMatch, focusing on model training in scenarios with few labeled samples.
Chain of Attack: On the Robustness of Vision-Language Models Against Transfer-Based Adversarial Attacks
Peng Xie (Hong Kong University of Science and Technology), Kani Chen (Hong Kong University of Science and Technology)
CodeAdversarial AttackTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Evaluate the robustness of VLM in black-box transfer adversarial attacks, propose the Chain of Attack (CoA) method to generate adversarial images, and introduce LLM-based ASR automatic evaluation metrics.
π― What it does: This paper proposes an unsupervised single image de-raining framework called CSUD, which eliminates the dependence on real rain-clean paired data through pseudo-pair generation and self-reconstruction strategies.
π― What it does: A new tokenization method called Charm is proposed, which retains the composition, high definition, aspect ratio, and multi-scale information of images without modifying the pre-trained ViT structure, in order to enhance the performance of image aesthetic assessment and image quality assessment.
Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment
Yang Bai (Wuhan University), Mang Ye (Wuhan University)
CodeRetrievalTransformerLarge Language ModelContrastive LearningImageTextMultimodality
π― What it does: A framework for dialogue-oriented person retrieval (ChatPR) is proposed and implemented, which refines the retrieval query through interactive multi-turn dialogue and constructs the first ChatPedes dataset using dialogue data generated by large language models;
Cheb-GR: Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification
Jinxi Yang (Wuhan University), Mang Ye (Wuhan University)
CodeRecognitionRetrievalGraph Neural NetworkImage
π― What it does: The Cheb-GR method is proposed, utilizing adaptive neighbor search guided by the Chebyshev theorem and graph convolution rearrangement, eliminating the reliance on k-nearest neighbor search in traditional rearrangement.
CheXwhatsApp: A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices
Mariamma Antony (Indian Institute of Science), Chiranjib Bhattacharyya (Indian Institute of Science)
CodeClassificationObject DetectionCompressionExplainability and InterpretabilityConvolutional Neural NetworkImage
π― What it does: The CheXwhatsApp dataset has been constructed and made publicly available, containing 141,804 pairs of original and WhatsApp-compressed chest X-ray images. Three quantification metrics, PIP, OLS, and LI, are proposed to evaluate the prediction, interpretability, and localization instability caused by compression.
CheXWorld: Exploring Image World Modeling for Radiograph Representation Learning
Yang Yue (Tsinghua University), Gao Huang (Tsinghua University)
CodeClassificationSegmentationRepresentation LearningTransformerContrastive LearningWorld ModelImageBiomedical Data
π― What it does: This paper proposes CheXWorld, a self-supervised world modeling framework for chest X-ray images, aimed at learning anatomical structures, layouts, and domain transformation knowledge.
CholecTrack20: A Multi-Perspective Tracking Dataset for Surgical Tools
Chinedu Innocent Nwoye (University of Strasbourg), Nicolas Padoy (University of Strasbourg)
CodeObject DetectionObject TrackingVideoBenchmark
π― What it does: Created and publicly released the CholecTrack20 multi-view surgical tool tracking dataset, which includes multi-class and multi-tool tracking annotations for 20 complete cholecystectomy surgery videos.
π― What it does: This paper reveals the existence of a 'leading silence' short-term silence bias in audio-video deepfake datasets and proposes an unsupervised learning framework called AVH-Align, which utilizes only real samples. It enhances detection robustness by aligning AV-HuBERT self-supervised features, avoiding reliance on dataset shortcuts like brief silences.
Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization
Anubhav Jain (New York University), Yuki Mitsufuji (Sony Group Corporation)
CodeGenerationDiffusion modelImage
π― What it does: This study investigates the phenomenon of memorization in diffusion models and proposes techniques to reduce memorization by detecting and avoiding attractor basins in the diffusion trajectory, using dynamic or static transition points and Opposite Guidance.
π― What it does: This paper proposes an unsupervised multi-label classification method called Classifier-guided CLIP Distillation (CCD), which utilizes pseudo-labels from CLIP and combines classifier-guided local views and debiasing techniques to train multi-label recognition models without manual annotation.
Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers
Quentin Guimard (University of Trento), Elisa Ricci (University of Trento)
CodeClassificationRetrievalLarge Language ModelImageTextRetrieval-Augmented Generation
π― What it does: An unsupervised framework named CLASSIFIER-TO-BIAS (C2B) is proposed, which can automatically identify the biases of visual classifiers solely based on the textual description of the given classification task and a pre-trained model.
ClearSight: Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models
Hao Yin (University of Science and Technology of China), Zilei Wang (University of Science and Technology of China)
CodeObject DetectionGenerationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
π― What it does: A training-free and tool-free visual enhancement fusion method (VAF) is proposed, which enhances the attention to visual signals in multimodal large language models by adjusting the mid-layer attention weights, thereby reducing the generation of false objects.
CLIP-driven Coarse-to-fine Semantic Guidance for Fine-grained Open-set Semi-supervised Learning
Xiaokun Li (Beijing Jiaotong University), Qingji Guan (Beijing Jiaotong University)
CodeClassificationRecognitionTransformerVision Language ModelContrastive LearningImage
π― What it does: A coarse-fine layered semantic guidance framework based on CLIP (CFSG-CLIP) is designed, which first filters local visual features related to categories through a semantic filtering module, and then further refines the features in the fine-grained branch through a visual semantic injection strategy to achieve efficient differentiation between fine-grained IDs and OOD samples.
CLOC: Contrastive Learning for Ordinal Classification with Multi-Margin N-pair Loss
Dileepa Pitawela (University of Adelaide), Hsiang-Ting Chen (University of Adelaide)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkContrastive LearningImage
π― What it does: This paper proposes an ordinal classification method CLOC based on multi-margin N-pair contrastive learning, which utilizes learnable margins to preserve category order and reduce critical decision errors.
π― What it does: Proposes the Implicit Motion-Audio Entanglement (IMAE) method, which generates realistic upper-body pose videos directly from audio without using additional inputs.
π― What it does: A dual-stage learning framework based on compression and adaptation (CoA) is proposed, which performs adaptation through dual-layer optimization in the real domain after compressing the model in the synthetic domain;
π― What it does: This paper addresses the issue of boundary ambiguity in 3D Gaussian Splatting (3DGS) for object segmentation and proposes a joint optimization method for semantics and texture segmentationβCOB-GS.
π― What it does: A multi-layer visual feature alignment framework called CocoER is proposed, which is based on competitive and collaborative mechanisms. It extracts head, body, and background features through cross-layer attention and generates pseudo-labels via a vocabulary information alignment module to guide feature refinement.
CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification
Wenlong Yu (Tianjin University), Qinghua Hu (Tianjin University)
CodeExplainability and InterpretabilityConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelImage
π― What it does: A deep visual model interpretability framework based on automated concept decoding, separation, ambiguity quantification, and chain-based explanation is proposed.
π― What it does: This paper proposes Collaborative Decoding (CoDe), which splits the coarse and fine scale generation tasks of visual autoregressive models into a large model responsible for low-frequency global sketches and a small model responsible for high-frequency details, significantly improving inference speed and memory utilization.
Ka Chun Shum (Hong Kong University of Science and Technology), Sai-Kit Yeung (Hong Kong University of Science and Technology)
CodeGenerationData SynthesisDiffusion modelImage
π― What it does: This paper proposes a method for achieving fine-grained color conditional generation in diffusion models, capable of generating images that are highly consistent with given color patterns (color values and proportions) while maintaining diversity and quality of content.
ComfyBench: Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems
Xiangyuan Xue (Shanghai Artificial Intelligence Laboratory), Lei Bai (Shanghai Artificial Intelligence Laboratory)
CodeLarge Language ModelAgentic AIBenchmarkRetrieval-Augmented Generation
π― What it does: A ComfyBench benchmark was constructed, and ComfyAgent was proposed to automatically design and generate ComfyUI workflows using LLM.
CoMM: A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation
Wei Chen (Hong Kong University of Science and Technology), Long Chen (Hong Kong University of Science and Technology)
CodeGenerationData SynthesisTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodalityBenchmark
π― What it does: We constructed and released CoMMβa high-quality, coherent interleaved text-image datasetβand proposed four new benchmarks for interleaved text-image generation and evaluation.
Complementary Advantages: Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising
Yuchen Wang (Beijing Institute of Technology), Hua Huang (Beijing Normal University)
CodeRestorationConvolutional Neural NetworkImage
π― What it does: This paper proposes an image denoising network called FCENet, which utilizes near-infrared (NIR) images to assist in achieving high-quality denoising of RGB images through frequency domain information fusion.
π― What it does: A combinatorial-based multi-label target universal perturbation generation framework is proposed, capable of synthesizing attack perturbations for any label combination in linear time;
π― What it does: A trainable and exchangeable angle matrix rotation position encoding (ComRoPE) is proposed, achieving scalability and robustness for Transformer position encoding.
CodeClassificationDomain AdaptationTransformerVision Language ModelImageMultimodality
π― What it does: This study investigates how to apply split conformal prediction to large-scale pre-trained vision-language models (such as CLIP) and proposes an unsupervised transfer learning framework based on optimal transport, named Conf-OT, to address the distribution shift between the source and target domains.
π― What it does: A posterior sampling method based on Langevin dynamics in the noise space of pre-trained generative models is proposed, which can efficiently generate diverse posterior samples.
π― What it does: This paper proposes Context-Aware Multimodal Pretraining (LIxP), which incorporates cross-attention contextual buffering into contrastive pretraining, enabling visual-text models to perform metric-based few-shot adaptation without additional training after training.
π― What it does: The ConText-CIR framework is proposed, which trains with Text Concept-Consistency loss and achieves efficient image-text combined retrieval through cross-modal attention.
Context-Enhanced Memory-Refined Transformer for Online Action Detection
Zhanzhong Pang (National University of Singapore), Angela Yao (National University of Singapore)
CodeRecognitionTransformerVideo
π― What it does: To address the frame representation learning bias caused by the inconsistency between training and inference in Online Action Detection (OAD), a new Transformer framework called Context-Enhanced Memory-Refined Transformer (CMeRT) is proposed.
Contextual AD Narration with Interleaved Multimodal Sequence
Hanlin Wang (Nanjing University), Limin Wang (Nanjing University)
CodeRecognitionGenerationTransformerLarge Language ModelContrastive LearningVideoTextMultimodalityAudio
π― What it does: A unified framework called Uni-AD is proposed, which automatically generates audio descriptions (AD) using interleaved multimodal sequences and pre-trained LLMs.
Continual SFT Matches Multimodal RLHF with Negative Supervision
Ke Zhu (Nanjing University), Jingdong Wang (Baidu)
CodeComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodality
π― What it does: Proposes a negative supervision fine-tuning (nSFT) method that transforms negative feedback from RLHF into SFT loss using a single model, thereby efficiently aligning visual language models.
π― What it does: A continuous crowd behavior generation framework called CrowdES is designed based on diffusion models and state-switching dynamics, capable of automatically generating diverse and realistic crowd trajectories continuously from a single scene image while supporting user interactive control.
π― What it does: A rewritten diffusion model DeCo-Diff is proposed, which only corrects abnormal regions while keeping normal areas unchanged in multi-class unsupervised anomaly detection.
π― What it does: The CoSDH framework is proposed, utilizing supply-demand awareness and late-stage mixed fusion to achieve efficient communication for collaborative 3D object detection.
Crab: A Unified Audio-Visual Scene Understanding Model with Explicit Cooperation
Henghui Du (Renmin University of China), Di Hu (Renmin University of China)
CodeRecognitionObject DetectionSegmentationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodalityAudio
π― What it does: A unified audio-video scene understanding model, Crab, is proposed, capable of simultaneously working on various audio-video tasks such as temporal localization, spatial localization, spatiotemporal reasoning, and pixel-level understanding.
Critic-V: VLM Critics Help Catch VLM Errors in Multimodal Reasoning
Di Zhang (Fudan University), Dongzhan Zhou (Shanghai Artificial Intelligence Laboratory)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelTextMultimodality
π― What it does: The Critic-V framework is proposed, which splits the reasoning process of visual language models into a Reasoner and an external Critic, iteratively improving the reasoning path through natural language criticism.
CroCoDL: Cross-device Collaborative Dataset for Localization
Hermann Blum (Lamarr Institute for AI and Security), Zuria Bauer (ETH Zurich)
CodePose EstimationRetrievalRobotic IntelligenceSimultaneous Localization and MappingMultimodalityBenchmark
π― What it does: The CroCoDL dataset has been constructed and released, providing multimodal visual localization data recorded across 10 different scenes and 4 types of devices (robots, handheld smartphones, MR headsets, NavVis scanners), and generating high-precision pseudo Ground-Truth based on LaMAR.
π― What it does: A cross-modal causal relationship alignment framework (CRA) is proposed, utilizing Gaussian Smoothing Grounding, Cross-Modal Alignment, and Explicit Causal Intervention to achieve weakly supervised video question answering and temporal localization.
π― What it does: This paper proposes a 3D multi-object discovery method DIOD-3D based on 2D motion information, and achieves mutual supervision between 2D and 3D through the cross-modal distillation framework xMOD, enabling unsupervised multi-modal object discovery.
π― What it does: This study first collected and annotated 21,930 pairs of PET-CT lung images, constructing the publicly available dataset PCLT20K; subsequently, a cross-modal interactive perception network (CIPA) based on the Mamba state space model was proposed for lung tumor segmentation in PET-CT images.
π― What it does: To address the problem of SAR image registration, a Cross Reject Open Set Registration method (CroR-OSIR) is proposed, which achieves accurate selection and localization of matching points through cross open set classification and supervised contrastive learning in two image domains.
π― What it does: A single-network semi-supervised breast ultrasound image segmentation framework CSC-PA is proposed, which enhances segmentation performance by utilizing cross-image semantic correlation.
Curriculum Coarse-to-Fine Selection for High-IPC Dataset Distillation
Yanda Chen (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
CodeKnowledge DistillationImage
π― What it does: This paper proposes a course-based coarse-to-fine selection method named CCFS for ensemble data distillation under high per-class sample (high-IPC) conditions, enhancing the distillation effect by gradually selecting suitable real images combined with synthetic images.
π― What it does: A new knowledge distillation framework called CustomKD is proposed, utilizing large-scale visual foundation models (such as DINOv2 and OpenCLIP) as teachers. Through alternating training of feature customization and distillation, it significantly enhances the performance of edge models (such as MobileNetV3 and WideResNet) in scenarios with unlabeled data (Unsupervised Domain Adaptation UDA and Semi-Supervised Learning SSL).
π― What it does: This paper proposes a Continuous Testing Time Adaptation (CTTA) framework DCTTA, designed for LiDAR semantic segmentation, which can online update the model in continuously changing environments.
DAMM-Diffusion: Learning Divergence-Aware Multi-Modal Diffusion Model for Nanoparticles Distribution Prediction
Junjie Zhou (Nanjing University of Aeronautics and Astronautics), Wei Shao (Nanjing University of Aeronautics and Astronautics)
CodeGenerationData SynthesisDiffusion modelMultimodalityBiomedical Data
π― What it does: A diffusion model called DAMM-Diffusion, which integrates single-modal and multi-modal branches, is proposed for predicting the distribution of nanoparticles under tumor microenvironment conditions (vascular and cell nuclei).
Daniel Feijoo (Cidaut AI), Marcos V. Conde (Computer Vision Lab)
CodeRestorationConvolutional Neural NetworkImage
π― What it does: A lightweight multi-task network called DarkIR is proposed for simultaneous denoising, deblurring, and low-light enhancement in nighttime/low-light environments.
Data Distributional Properties As Inductive Bias for Systematic Generalization
Felipe del Rio (Pontificia Universidad Catolica de Chile), Alvaro Soto (Pontificia Universidad Catolica de Chile)
CodeTransformerMultimodality
π― What it does: Exploring how the characteristics of training data distribution (diversity, burstiness, latent intervention) enhance the system generalization of multimodal models.
π― What it does: This paper proposes a diffusion model named MorphFace, which can generate identity-preserving and stylistically diverse synthetic faces to enhance the diversity of training data for facial recognition models.
Debiasing Multimodal Large Language Models via Noise-Aware Preference Optimization
Zefeng Zhang (Institute of Information Engineering, Chinese Academy of Sciences), Tingwen Liu (Institute of Information Engineering, Chinese Academy of Sciences)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningMultimodalityBenchmark
π― What it does: This paper proposes a multimodal large language model debiasing method based on preference optimization, first constructing the RLAIF-V-Bias dataset and designing the NaPO algorithm to suppress modal bias.
π― What it does: The DeCLIP framework is proposed, which decouples the self-attention module of CLIP to separately learn local identification ability and spatial consistency, enhancing the dense prediction performance of open vocabulary.
π― What it does: This paper addresses a black-box watermarking method for image-to-image models, revealing the vulnerability of the decoder to gradient attacks, and proposes the Decoder Gradient Shield (DGS) mechanism, which redirects and scales gradients in a black-box API to prevent attackers from training a de-watermarking network.
π― What it does: A decoder-based reverse mapping method called DeDe is proposed for detecting hidden backdoor samples in unsupervised self-supervised learning (SSL) encoders.
π― What it does: A UAVTC dataset for long-term fine-grained tree change detection is proposed, and a hypercurvature Siamese network (HSN) is designed to capture the hierarchical structure of tree changes.
π― What it does: Designed and implemented DeepCompress-ViT, which significantly compresses the Vision Transformer using an encoder-decoder structure, and ensures accuracy is not lost through optimized inference-time decoding.
π― What it does: A continuous representation based on Deformable Centerline (DeformCL) is proposed for the extraction and segmentation of blood vessels in 3D medical images;
π― What it does: A global unified model DFPIR based on degradation-aware feature perturbation is designed, capable of handling multiple degradation tasks such as image denoising, dehazing, deraining, deblurring, and low-light enhancement in one go.
DejaVid: Encoder-Agnostic Learned Temporal Matching for Video Classification
Darryl Ho (Massachusetts Institute of Technology), Samuel Madden (Massachusetts Institute of Technology)
CodeClassificationTransformerVideo
π― What it does: An Encoder-agnostic DejaVid method is proposed, which encodes videos into variable-length time series embeddings (TSE) and uses Dynamic Time Warping (DTW) with learnable temporal feature weights to enhance video classification performance.
π― What it does: Proposes the EarlyLate training strategy, which optimizes synthetic samples in stages to enhance the diversity and efficiency of dataset distillation under batch-to-global matching.
π― What it does: A depth-guided beam sampling strategy is proposed, which groups adjacent rays for joint sampling and adaptively controls the number of sampling points to improve the rendering speed and quality of general NeRF.
Derivative-Free Diffusion Manifold-Constrained Gradient for Unified XAI
Won Jun Kim (KAIST), Jong Chul Ye (KAIST)
CodeExplainability and InterpretabilityDiffusion modelImageBiomedical Data
π― What it does: This paper proposes a FreeMCG method based on diffusion models and EnKF for the unified implementation of black-box feature importance explanation and counterfactual explanation.
π― What it does: This paper proposes DeSiRe-GS, a self-supervised static-dynamic decomposition and surface reconstruction framework based on 4D Gaussian splatting.
Detect Any Mirrors: Boosting Learning Reliability on Large-Scale Unlabeled Data with an Iterative Data Engine
Zhaohu Xing (Hong Kong University of Science and Technology), Lei Zhu (Hong Kong University of Science and Technology)
CodeObject DetectionSegmentationAnomaly DetectionKnowledge DistillationTransformerLarge Language ModelImageVideoMultimodality
π― What it does: By constructing a semi-supervised mirror detection framework and collecting approximately 400,000 unlabeled mirror images, the quality of pseudo-labels is improved using an iterative data engine, achieving more robust mirror detection.
Detect-and-Guide: Self-regulation of Diffusion Models for Safe Text-to-Image Generation via Guideline Token Optimization
Feifei Li (Fudan University), Min Yang (Fudan University)
CodeGenerationDiffusion modelImageText
π― What it does: This paper proposes the Detect-and-Guide (DAG) framework, which performs self-inspection in text-to-image diffusion models by optimizing prompt words and extracting Class Activation Maps (CAM). It then employs adaptive safety guidance during the sampling process to finely eliminate detected unsafe areas, achieving safe generation of sexual content.
π― What it does: Proposes the Perturbation Forgery method, which perturbs the noise distribution of known attacks, generates sparse masks, and synthesizes pseudo-adversarial samples to train a binary classifier for detecting unknown adversarial attacks.
Detecting Backdoor Attacks in Federated Learning via Direction Alignment Inspection
Jiahao Xu (University of Nevada), Rui Hu (University of Nevada)
CodeAnomaly DetectionFederated LearningImage
π― What it does: A backdoor defense method based on directional alignment detection, AlignIns, is proposed to identify and filter malicious model updates in federated learning.
π― What it does: This paper proposes the PASS method, which automatically selects and optimizes a small subset of attributes from a large attribute pool through partial optimal transport and curriculum learning, to achieve simultaneous detection of known and unknown objects in open-world object detection (OWOD).
π― What it does: This paper proposes a post-hoc OOD detection method based on the theory of Neural Collapse, utilizing the alignment of ID sample features with the last layer weight vectors and the L1 norm of the features to distinguish OOD samples.
π― What it does: A joint, detection-friendly end-to-end framework called UniCD is proposed for simultaneously performing non-uniformity correction and target detection of infrared drone images under low-frequency non-uniformity conditions.
π― What it does: A dual-approximation denoising Brownian bridge model (Dualβapprox Bridge) is proposed, achieving deterministic image-to-image translation.
Devils in Middle Layers of Large Vision-Language Models: Interpreting, Detecting and Mitigating Object Hallucinations via Attention Lens
Zhangqi Jiang (National University of Defense Technology), Xu Yang (Southeast University)
CodeObject DetectionGenerationExplainability and InterpretabilityTransformerVision Language ModelImage
π― What it does: This paper conducts an in-depth study of the object hallucination mechanism in large visual-language models (LVLM) from the perspective of attention and proposes a no-training-cost hallucination suppression method based on intermediate layer attention modulation.
π― What it does: DFormerv2 is proposed, a RGB-D semantic segmentation model that directly guides self-attention using the geometric priors of depth maps, without the need for an additional depth encoder, significantly improving segmentation performance.
π― What it does: A diffusion model DiC based on pure 3x3 convolutions is proposed, achieving image generation under the Encoder-Decoder hourglass architecture.
π― What it does: A Diff-Palm model based on polynomial line drawing and K-step noise sharing sampling is proposed to generate high-quality, controllable variations of palm print images on a large scale, and to directly train recognition networks without fine-tuning on real data.
π― What it does: A LiDAR pose estimation network called DiffLO is proposed, which integrates semantic awareness and diffusion models to improve the accuracy and robustness of large-scale LiDAR pose estimation.
π― What it does: Proposed the Diffusion-4K framework, achieving direct generation of 4K level text-to-image, and constructed the Aesthetic-4K evaluation benchmark.
π― What it does: This paper proposes an event-guided image deblurring framework (EGDeblurring) based on a diffusion model, which can perform deblurring using only blurred images.
π― What it does: An end-to-end multi-view method based on a denoising diffusion model is proposed, which directly predicts pixel-level ray origins and endpoints, thereby obtaining both scene geometry and camera poses simultaneously.
π― What it does: This paper proposes DIFFVSGG, a method for generating online video scene graphs based on latent diffusion models, which can update object categories, bounding boxes, and relationships in real-time for each frame, supporting continuous temporal reasoning.
π― What it does: This paper proposes a diffusion model-based infrared image super-resolution method called DifIISR, which achieves high-quality infrared image reconstruction by injecting visual and perceptual gradients during the reverse diffusion process.
π― What it does: This paper proposes a complete pipeline called DIFIX3D+ that utilizes a single-step diffusion model (DIFIX) to improve NeRF and 3D Gaussian Splatting for reconstruction and novel view synthesis.
DiGIT: Multi-Dilated Gated Encoder and Central-Adjacent Region Integrated Decoder for Temporal Action Detection Transformer
Ho-Joong Kim (Korea University), Seong-Whan Lee (Korea University)
CodeRecognitionObject DetectionTransformerVideo
π― What it does: A time action detection framework based on Transformer, called DiGIT, is proposed, which includes a multi-dilated gated encoder and a center-adjacent region integrated decoder.
π― What it does: A Digital Twin Catalog (DTC) dataset has been proposed and released, containing high-precision 3D digital twin models of 2000 real objects, with millimeter-level geometric accuracy, photorealistic PBR materials, and multi-view evaluation data collected using DSLR and AR glasses.
Dispider: Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction
Rui Qian (Chinese University of Hong Kong), Jiaqi Wang (Shanghai Innovation Institute)
CodeGenerationOptimizationTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
π― What it does: Designed and implemented Dispider, a decoupled framework for real-time interactive video large language models, supporting continuous monitoring, timely decision-making, and asynchronous responses;
π― What it does: This paper proposes the Dataset Distillation for Long-Tail Data (LTDD) task, which can compress imbalanced raw data into a small-scale, uniformly distributed synthetic dataset.