🎯 What it does: Propose a fuzzy knowledge distillation (BKD) technique, mapping randomly blurred SNN features through a recovery block to continuous features, thereby more effectively distilling features from the teacher ANN, and combining with logits distillation to enhance the performance of learning-based spiking neural networks (SNNs).
🎯 What it does: Propose a blind image deconvolution framework that utilizes generative adversarial networks (GANs) to learn the distribution of blur kernels and obtain kernel initialization through latent space mapping, followed by jointly optimizing the image and kernel via a deep network within the latent kernel space.
🎯 What it does: This paper proposes an end-to-end Gaze Object Segmentation (GOS) framework that generates pixel-level masks using a visual foundation model (SAM), and accurately predicts pixel-level masks of human-gazed objects through a unified detection and segmentation Transformer, head feature reconstruction, and a spatial-to-target gaze regression strategy.
Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory
Sensen Gao (Nankai University), Qing Guo (Agency for Science, Technology and Research)
CodeAdversarial AttackVision Language ModelMultimodality
🎯 What it does: Propose a novel method to enhance the transferability of adversarial attacks in vision-language pre-training models: by sampling from the cross-region of adversarial trajectories and combining text-guided sample selection to generate diverse and more transferable multimodal adversarial examples; simultaneously, introduce a cross-region offset strategy in the text modality to further reduce the risk of overfitting on the target model.
🎯 What it does: This paper proposes Brain-ID, a contrast-agnostic anatomical feature learning framework based on synthetic data generation, for unified representation in brain imaging.
BRIDGE: Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues
Sara Sarto (University of Modena and Reggio Emilia), Rita Cucchiara (University of Modena and Reggio Emilia)
CodeRecognitionTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposed a new, reference-free image caption evaluation metric called BRIDGE, which integrates fine-grained visual features into multimodal pseudo-descriptions to more accurately measure the correspondence between generated text and images.
🎯 What it does: Propose the LaVi-Bridge framework, achieving seamless integration of any pre-trained language model with a generative vision model for text-to-image diffusion generation tasks.
🎯 What it does: Propose a dual-branch, plug-and-play image inpainting model called BrushNet, which can hierarchically inject pixel-level occlusion information into pre-trained diffusion models, achieving high-quality, semantically consistent image restoration.
BurstM: Deep Burst Multi-scale SR using Fourier Space with Optical Flow
EungGu Kang (Daegu Gyeongbuk Institute of Science and Technology), Kyong Hwan Jin (Korea University)
CodeSuper ResolutionOptical FlowImage
🎯 What it does: Proposed a multi-frame super-resolution network named BurstM, which can precisely align low-resolution burst images without using DCN by leveraging optical flow, and reconstructs high-frequency textures by predicting continuous Fourier coefficients in the Fourier space; simultaneously supports multi-scale (×2, ×3, ×4) super-resolution under a single model;
🎯 What it does: Propose the Zero-Shot Compositional Action Recognition (ZS-CAR) task, construct the corresponding Sth-com benchmark dataset, and develop the Component-to-Composition (C2C) learning framework to accomplish this task.
🎯 What it does: Proposed the first publicly available RGB-thermal imaging dataset aimed at supporting aerial robots operating in natural environments. The dataset contains synchronized RGB, thermal imaging, global positioning, and inertial data, along with semantic segmentation annotations for 10 common categories.
🎯 What it does: Propose the first camera-radar cross-modal gait recognition framework CL-Gait, which extracts 2D gait silhouettes and 3D point cloud features using dual-stream networks, and achieves cross-modal matching through alignment learning.
Can OOD Object Detectors Learn from Foundation Models?
Jiahui Liu (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)
CodeObject DetectionData SynthesisAnomaly DetectionConvolutional Neural NetworkLarge Language ModelDiffusion modelImage
🎯 What it does: This paper leverages large language models, text-to-image diffusion models, and segmentation models to automatically generate and annotate scene-level synthetic OOD samples, aiming to enhance OOD object detection performance.
Can Textual Semantics Mitigate Sounding Object Segmentation Preference?
Yaoting Wang (Renmin University of China), Di Hu (Renmin University of China)
CodeSegmentationTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityAudio
🎯 What it does: Proposes the TeSO method, which leverages text semantics to enhance audio-visual segmentation through modules such as image description, frozen LLM inference of potential sound objects, semantic-driven audio modeling (SeDAM), and masked query prompting (PMQS).
Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning
Ali Cheraghian (Data61, CSIRO), Mehrtash Harandi (Monash University)
CodeClassificationTransformerLarge Language ModelPrompt EngineeringVision Language ModelTextPoint Cloud
🎯 What it does: This paper proposes a complete framework called C3PR for few-shot incremental learning (FSCIL) of 3D point clouds using the CLIP model.
🎯 What it does: Propose CanonicalFusion, which simultaneously predicts depth and compressed linear blend skinning (LBS) weights using multiple images, and generates a drivable 3D human model based on this;
🎯 What it does: Proposed CARB-Net, which integrates camera angular precision with radar depth perception, achieving breakthroughs in addressing the challenge of radar's inability to distinguish nearby targets when angular resolution is insufficient.
🎯 What it does: Propose CardiacNet, a reconstruction-based deep learning framework for learning local structural and motion abnormalities from cardiac ultrasound videos and assessing cardiac diseases.
Cascade Prompt Learning for Visual-Language Model Adaptation
Ge Wu (Nankai University), Xiang Li (Nankai University)
CodeClassificationDomain AdaptationKnowledge DistillationPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: Propose a two-stage Prompt learning framework called CasPL, which enhances domain-agnostic knowledge and adapts to task-specific knowledge on vision-language models (e.g., CLIP), and seamlessly integrates as a plugin into existing Prompt methods.
🎯 What it does: Proposed a multi-instance learning framework based on cascading Dirichlet process (cDP-MIL), which clusters WSI local patches using deep variational parameter DP, generates sliding window-level representations, and provides uncertainty estimation and tumor localization;
Zhengyuan Jiang (Duke University), Neil Zhenqiang Gong (Pennsylvania State University)
CodeAdversarial AttackImage
🎯 What it does: This paper proposes an image watermarking framework based on randomized smoothing, which provides provably robust watermarking against removal and forgery attacks under L2 norm constraints, and presents three smoothing schemes (multi-class, multi-label, regression) along with derivations and estimations of their robustness lower/upper bounds.
Challenging Forgets: Unveiling the Worst-Case Forget Sets in Machine Unlearning
Chongyu Fan (Michigan State University), Sijia Liu (Michigan State University)
CodeOptimizationSafty and PrivacyImage
🎯 What it does: Proposes identifying the most challenging forgetting set (worst-case forget set) from an adversarial perspective to evaluate the robustness of machine unlearning methods.
🎯 What it does: Designed and trained a universal model called Chameleon that can adapt to any dense visual prediction task under extremely few samples;
ChEX: Interactive Localization and Region Description in Chest X-rays
Philip Müller (Technical University of Munich), Daniel Rueckert (Technical University of Munich)
CodeObject DetectionGenerationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBiomedical Data
🎯 What it does: Propose a multi-task model named ChEX that can generate and locate textual descriptions of lesions or structures in chest X-rays based on text prompts or bounding box queries.
CIC-BART-SSA: : Controllable Image Captioning with Structured Semantic Augmentation
Kalliopi Basioti (Rutgers University), Afsaneh Fazly (Samsung AI Centre - Toronto)
CodeGenerationTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Developed a structured semantic enhancement (SSA) method based on AMR, automatically generating diverse focused image captions and using them to train the controlled image captioning model CIC-BART-SSA.
CipherDM: Secure Three-Party Inference for Diffusion Model Sampling
Xin Zhao (University of Chinese Academy of Sciences), Zhendong Zhao (Chinese Academy of Sciences)
CodeGenerationSafty and PrivacyDiffusion modelImage
🎯 What it does: Designed and implemented CipherDM, a three-party secure multi-party computation (3PC) framework for secure sampling under diffusion models (DDPM, DDIM, Stable Diffusion), ensuring privacy of model parameters and user inputs.
🎯 What it does: Propose a data-free post-training quantization method called CLAMP-ViT, specifically designed for unified quantization of weights and activations in visual Transformers (ViT), adaptively generating semantically rich synthetic data during the process to achieve high-precision compression.
CLAP: Isolating Content from Style through Contrastive Learning with Augmented Prompts
Yichao Cai (University of Adelaide), Javen Qinfeng Shi (University of Adelaide)
CodeDomain AdaptationRepresentation LearningData-Centric LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: By performing contrastive learning with image and text augmentation on a pre-trained CLIP model, content features are extracted and style information is removed, enhancing the generalization ability of multi-modal representations.
Class-Incremental Learning with CLIP: Adaptive Representation Adjustment and Parameter Fusion
Linlan Huang (Nankai University), Xialei Liu (Nankai University)
CodeRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes the RAPF method, achieving class-incremental learning on the CLIP model through text-guided neighborhood class separation and parameter fusion, significantly reducing catastrophic forgetting.
ClearCLIP: Decomposing CLIP Representations for Dense Vision-Language Inference
Mengcheng Lan (S-Lab Nanyang Technological University), Wayne Zhang (SenseTime Research)
CodeSegmentationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: Investigated the root causes of noise generated by CLIP in open-source semantic segmentation tasks, proposing ClearCLIP which significantly improves segmentation quality by removing residual connections, adopting self-attention mechanisms, and eliminating FFN.
CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection
Wuyang Li (Chinese University of Hong Kong), Yixuan Yuan (Chinese University of Hong Kong)
CodeObject DetectionVision Language ModelDiffusion modelAuto EncoderImageTextMultimodality
🎯 What it does: This paper proposes a continuous latent diffusion framework named CLIFF for open-vocabulary object detection (OVD), which enhances the detection performance of novel categories by enabling continuous distribution transfer between object, image, and text latent spaces.
🎯 What it does: Developed a Camera-based Respiratory Measurement model called CliffPhys based on Clifford neural networks, which utilizes optical flow and monocular depth estimation to obtain 2D vector fields and scalar fields, and captures geometric relationships through Clifford Neural Layers to achieve non-contact breathing rate measurement.
🎯 What it does: This paper proposes a generative network called CGNC that leverages CLIP text knowledge through a cross-attention module to achieve multi-target transferable directed adversarial attacks;
CloudFixer: Test-Time Adaptation for 3D Point Clouds via Diffusion-Guided Geometric Transformation
Hajin Shim (Korea Advanced Institute of Science and Technology), Eunho Yang (Korea Advanced Institute of Science and Technology)
CodeDomain AdaptationDiffusion modelPoint Cloud
🎯 What it does: Implement test-time adaptation (TTA) for 3D point cloud inference in distribution drift scenarios by enabling input adaptation during testing through geometric transformations guided by a pre-trained diffusion model, significantly enhancing model robustness.
🎯 What it does: Propose a new GAN training paradigm called CLR-GAN, treating the generator and discriminator as inverse processes. Introduce a latent space reconstruction task in the discriminator and a real image reconstruction task in the generator, forming latent space consistency loss and reconstruction loss to enhance GAN training stability and generation quality.
🎯 What it does: This paper proposes a cross-mechanism domain adaptation dataset named CMD, and builds a three-stage domain adaptation baseline method called DIG based on this dataset, aiming to address the performance degradation in point cloud detection across different LiDAR, radar, and camera sensors.
🎯 What it does: Proposes a video deblurring method (CMTA) that leverages the high temporal resolution information of event cameras, comprising two major modules: Cross-Modal Recursive Internal Feature Enhancement (CRIFE) and Event-Guided Cascaded Cross-Frame Temporal Feature Alignment (ECITFA), and constructs a novel real-world hybrid camera dataset named EVRB.
🎯 What it does: Proposed the Co-Student framework, which utilizes teacher-student dual students to collaboratively generate and denoise pseudo labels, thereby fully leveraging unlabeled objects in sparse annotation object detection (SAOD) tasks.
🎯 What it does: This paper proposes a coarse-to-fine implicit surface learning framework that progressively constructs the SDF of hand-object interactions using RGB-D inputs;
CoDA: Instructive Chain-of-Domain Adaptation with Severity-Aware Visual Prompt Tuning
ZiYang Gong, Zhenming Ji (Sun Yat-sen University)
CodeSegmentationDomain AdaptationAutonomous DrivingPrompt EngineeringVision Language ModelImageChain-of-Thought
🎯 What it does: Propose the CoDA method, combining Chain-of-Domain phased adaptation and Severity-Aware Visual Prompt Tuning to achieve unsupervised domain adaptation for adverse scenarios;
🎯 What it does: Propose a graph-based masked autoencoder (GMAE) and priority scheduling method to generate large-scale city-level 2.5D layouts, considering multi-level context sensitivity.
CoLA: Conditional Dropout and Language-driven Robust Dual-modal Salient Object Detection
Shuang Hao (Huazhong University of Science and Technology), He Tang (Huazhong University of Science and Technology)
CodeObject DetectionConvolutional Neural NetworkPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Designed and implemented the CoLA framework to simultaneously address noise input and modality missing issues in dual-modal salient object detection, incorporating two modules: Language-Driven Quality Assessment (LQA) and Conditional Dropout (CD).
🎯 What it does: Propose the CoLeaF framework for weakly supervised audio-visual video parsing (AVVP), optimizing cross-modal context fusion in the embedding space through contrastive and collaborative learning, specifically distinguishing alignable and non-alignable events.
COM Kitchens: An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark
Atsushi Hashimoto (OMRON SINIC X Corp.), Yoshitaka Ushiku (Cookpad Inc.)
CodeRetrievalTransformerVision Language ModelVideoTextBenchmark
🎯 What it does: This paper collects and constructs the COM Kitchens dataset, which includes 40 hours of 145 uncut high-angle cooking videos, and provides fine-grained visual action maps along with corresponding text instructions.
🎯 What it does: Propose a dual-modal framework based on generative adversarial networks (GANs) that leverages facial priors and geometric symmetry priors to correct facial distortions and background line warping in wide-angle portraits.
Yue Zhang (Michigan State University), Gaurav Bharaj (Reality Defender Inc)
CodeAnomaly DetectionExplainability and InterpretabilityTransformerVision Language ModelContrastive LearningImageVideoMultimodality
🎯 What it does: Convert the deepfake detection task into a visual question answering (DD-VQA) task, construct a corresponding dataset, and train a BLIP-based multimodal Transformer model to generate authenticity judgments and textual explanations.
Fitim Abdullahu (Zurich University of Applied Sciences), Helmut Grabner (Zurich University of Applied Sciences)
CodeRetrievalRecommendation SystemTransformerVision Language ModelContrastive LearningImage
🎯 What it does: Built a 'shared interest' image evaluation framework based on Flickr user likes behavior, using this definition to mine common and subjective interest images.
🎯 What it does: This paper proposes CompGS, a method that significantly compresses model storage and inference costs while maintaining the real-time rendering and high quality of 3D Gaussian Splatting (3DGS).
Compositional Substitutivity of Visual Reasoning for Visual Question Answering
Chuanhao Li (Beijing Institute of Technology), Yunde Jia (Beijing Institute of Technology)
CodeTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: This paper investigates compositional substitutability in visual question answering and proposes a training framework based on a support set, enabling the model to recognize and maintain invariant representations for synonymous primitives (words, visual entities, references).
Comprehensive Attribution: Inherently Explainable Vision Model with Feature Detector
Xianren Zhang (Pennsylvania State University), Suhang Wang (Pennsylvania State University)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: Proposed a novel inherently interpretable visual model called COMET, which utilizes a selector to generate attribution maps and collaborates with a predictor, while incorporating a pre-trained feature detector to enhance attribution completeness and prevent interlocking between the selector and predictor.
ConceptExpress: Harnessing Diffusion Models for Single-image Unsupervised Concept Extraction
Shaozhe Hao (University of Hong Kong), Kwan-Yee K. Wong (University of Hong Kong)
CodeRecognitionTransformerDiffusion modelImage
🎯 What it does: This paper proposes a method that automatically extracts and reproduces individual concepts from a single multi-concept image under unlabeled conditions using a pre-trained diffusion model.
🎯 What it does: Proposed the CONDA framework, which leverages deep networks to transform raw pixel-level associations into deep association features, achieving high-quality co-occurrent object detection through evolutionary association generation, association condensation, and object-aware cycle-consistent loss.
🎯 What it does: To address the challenge of partial labels in multi-label class incremental learning, the Confidence Self-Calibration (CSC) framework is proposed, which mitigates false positives and catastrophic forgetting caused by overconfidence through calibrating label relationships and multi-label confidence.
🎯 What it does: Building upon existing 3D generation frameworks based on Score Distillation Sampling (SDS), the authors propose Guided Consistency Sampling (GCS) and Brightness-Equalized Generation (BEG), which significantly enhance the detail and realism of text-driven 3D assets by integrating the concept of consistency distillation.
Xulong Bai (Chinese Academy of Sciences), Shuhan Shen (Chinese Academy of Sciences)
CodeOptimizationImage
🎯 What it does: Proposes a complete system for reconstructing 3D line segments and their line trajectories from multi-view known camera poses, emphasizing consistency and completeness.
🎯 What it does: This paper proposes a continual learning framework ADPP for remote optical pulse (rPPG) measurement, addressing the issues of catastrophic forgetting and domain shift during multi-domain sequence learning.
Contourlet Residual for Prompt Learning Enhanced Infrared Image Super-Resolution
Xingyuan Li, Risheng Liu (Dalian University of Technology)
CodeSuper ResolutionConvolutional Neural NetworkTransformerPrompt EngineeringVision Language ModelImageText
🎯 What it does: Propose an infrared image super-resolution network combining Contourlet residual and prompt learning, which extracts high-frequency details through multi-scale and multi-directional Contourlet decomposition and achieves semantic-level optimization by guiding the model with positive and negative text prompts.
🎯 What it does: This study proposes CoDE, a contrastive learning-based deep fake embedding space, specifically designed to distinguish real images from forged images generated by diffusion models, and creates the D3 dataset containing 9.2M images.
🎯 What it does: Propose the CLSP method, which generates hard positive samples using an unsupervised diffusion model and incorporates them into a contrastive learning framework
Yuzhong Zhao (University of Chinese Academy of Sciences), Fang Wan (University of Chinese Academy of Sciences)
CodeGenerationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Propose a controllable region-level image description method called ControlCap, which reduces description degradation and enables diverse, customizable region descriptions through control words.
Controllable Contextualized Image Captioning: Directing the Visual Narrative through User-Defined Highlights
Shunqi Mao (University of Sydney), Weidong Cai (University of Sydney)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the Ctrl-CIC task and design two controllers (Prompting-based Controller and Recalibration-based Controller) to generate contextualized image descriptions based on user-specified highlights.
Controllable Navigation Instruction Generation with Chain of Thought Prompting
Xianghao Kong, Si Liu
CodeGenerationAutonomous DrivingTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Proposes C-Instructor, which utilizes large language models combined with path and visual information to achieve controllable and interpretable navigation instruction generation.
ControlLLM: Augment Language Models with Tools by Searching on Graphs
Zhaoyang Liu (Hong Kong University of Science and Technology), Wenhai Wang (Chinese University of Hong Kong)
CodeGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the ControlLLM framework, which helps large language models (LLMs) accurately invoke multimodal tools to complete complex tasks through a three-stage process (task decomposition, graph-based thinking search, and execution engine).
CoPT: Unsupervised Domain Adaptive Segmentation using Domain-Agnostic Text Embeddings
Cristina Mata (Stony Brook University), Michael S Ryoo (Stony Brook University)
CodeSegmentationDomain AdaptationLarge Language ModelVision Language ModelImageText
🎯 What it does: Propose CoPT, achieving unsupervised domain adaptation for semantic segmentation by leveraging covariance consistency loss on domain-agnostic text embeddings.
🎯 What it does: Propose the CoSIGN framework, which combines a pre-trained Consistency Model with ControlNet to solve various inverse problems (super-resolution, inpainting, deblurring, sparse-view CT reconstruction) within 1-2 sampling steps (or with multi-step refinement).
🎯 What it does: In unlabelled continual self-supervised learning, this paper proposes the CroMo-Mixup framework, which enhances the model's ability to retain knowledge of previous tasks and distinguish task IDs through cross-task data mixing and cross-model feature mixing.
Cross-Domain Semantic Segmentation on Inconsistent Taxonomy using VLMs
Jeongkee Lim (Sungkyunkwan University), Yusung Kim (Sungkyunkwan University)
CodeSegmentationDomain AdaptationVision Language ModelImage
🎯 What it does: Proposes the CSI method, achieving semantic segmentation for cross-domain inconsistent classification vocabularies in unsupervised domain adaptation through vision-language models.
CodeRecognitionDomain AdaptationTransformerPrompt EngineeringVision Language ModelImageVideoBenchmark
🎯 What it does: Constructed the cross-platform video person re-identification benchmark dataset G2AVReID, and proposed the cross-platform video re-identification method VSLA-CLIP based on CLIP.
🎯 What it does: Designed the Panorama-BEV Co-Retrieval Network, combining panoramic street views and explicit bird's-eye-view dual branches to achieve cross-view retrieval.
🎯 What it does: This paper proposes Crowd-SAM, a framework that utilizes SAM as an intelligent annotator for one-class object detection with few samples in crowded scenes;
🎯 What it does: Proposes the Cross-scan Object Transfer (CSOT) method, which uses HotspotNet to predict suitable placement positions, copying and pasting annotated objects into unannotated LiDAR scans to generate sparsely annotated training data, thereby achieving semi-supervised LiDAR object detection.
🎯 What it does: Propose the DCI framework and ICO refinement strategy to address the fidelity-editability trade-off in text-to-image model customization
🎯 What it does: This paper systematically reconstructs pose-based gait recognition, proposing to directly use heatmaps rather than skeletons as input, improving preprocessing and heatmap alignment processes, and constructing a global-local network with multi-stage fusion branches, significantly enhancing recognition performance and cross-dataset generalization capabilities.
CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction
Zhangchen Ye (Tsinghua University), Hang Zhao (Tsinghua University)
CodeAutonomous DrivingConvolutional Neural NetworkTransformerImageTime Series
🎯 What it does: Propose a CVT-Occ module that constructs a 3D cost volume based on gaze sampling and historical frame projection for 3D occupancy prediction in multi-view time series.
🎯 What it does: Proposed and made publicly available a large-scale event camera action recognition benchmark dataset named DailyDVS-200, and conducted systematic evaluations of multiple models on it.
🎯 What it does: Proposed DAMSDet, a Transformer framework for infrared-visible light fusion object detection, addressing the problems of multi-modal information competition and alignment.
🎯 What it does: Generate new images with predictable saliency changes by performing controllable photometric and semantic editing on images through saliency-guided cross-attention in the latent space of Stable Diffusion, and use these images as data augmentation to enhance the training set of deep saliency prediction models.
Data Overfitting for On-Device Super-Resolution with Dynamic Algorithm and Compiler Co-Design
Gen Li (Clemson University), Xiaolong Ma (Clemson University)
CodeSuper ResolutionVideo
🎯 What it does: This paper proposes a "Dy-DCA" framework that transfers the video block super-resolution task into a single dynamic deep neural network, combined with a content-aware data preprocessing pipeline, thereby significantly reducing model switching overhead.
🎯 What it does: This paper proposes a data-to-model distillation framework, D2M, which compresses knowledge from large real-world datasets into the learnable parameters of a pre-trained generative model, thereby generating small yet information-rich synthetic images for efficient training.
🎯 What it does: This paper proposes DataDream, a few-shot data generation framework based on LoRA fine-tuning of Stable Diffusion, and further fine-tunes CLIP with LoRA, using a small number of real images to generate a high-quality synthetic training set, significantly enhancing classification performance.
🎯 What it does: Propose an adaptive training trajectory (ATT) method that dynamically selects the optimal trajectory length for long-range dataset distillation, generating a more representative synthetic dataset.
Ziheng Qin (National University of Singapore), Yang You (National University of Singapore)
CodeRetrievalComputational EfficiencyData-Centric LearningVision Language ModelMultimodality
🎯 What it does: Propose InfoGrowth, an online data cleaning and incremental sampling framework that evaluates the noise and redundancy of new samples in a multi-modal embedding space, automatically filtering and incrementally building a high-quality, diverse dataset.
🎯 What it does: Propose an adaptive sampling dataset quantization method (DQAS) based on active learning, improving the traditional DQ process and achieving more precise category sample allocation.
🎯 What it does: Proposed a new fast sampling method, DC-Solver, which utilizes dynamic compensation (DC) to mitigate alignment errors in predictor-corrector sampling, and can rapidly predict compensation ratios through cascading polynomial regression, achieving efficient DPM sampling.
DCDM: Diffusion-Conditioned-Diffusion Model for Scene Text Image Super-Resolution
Shrey Singh (Indian Institute of Technology), Partha Pratim Roy (Indian Institute of Technology)
CodeSuper ResolutionVision Language ModelDiffusion modelImage
🎯 What it does: To address the super-resolution problem of scene text images, a dual diffusion model named DCDM is proposed, which uses low-resolution images and character-level text embeddings as conditions for super-resolution.
🎯 What it does: Propose a causal intervention framework based on feature separation (FSCI) to achieve cross-domain generalization for gaze estimation without target domain data.
DEAL: Disentangle and Localize Concept-level Explanations for VLMs
Tang Li (University of Delaware), Xi Peng (University of Delaware)
CodeExplainability and InterpretabilityLarge Language ModelVision Language ModelContrastive LearningImageChain-of-Thought
🎯 What it does: Proposed an unsupervised method called DEAL that leverages large language models to generate discriminative visual concepts, and enhances the interpretability and predictive accuracy of Vision-Language Models for fine-grained concepts through contrastive learning, decoupling, and localization constraints.
DECIDER: Leveraging Foundation Model Priors for Improved Model Failure Detection and Explanation
Rakshith Subramanyam (Axio.ai), Jayaraman J. Thiagarajan (Lawrence Livermore National Laboratory)
CodeAnomaly DetectionExplainability and InterpretabilityConvolutional Neural NetworkLarge Language ModelVision Language ModelImageText
🎯 What it does: This paper proposes the DECIDER method, which generates task-related core attributes using a large language model (LLM) and constructs a 'bias-aware' classifier PIM with a vision-language model (VLM). It detects model failure by analyzing the prediction differences between the original classifier and PIM, and provides interpretable failure reasons through attribute ablation.
Deciphering the Role of Representation Disentanglement: Investigating Compositional Generalization in CLIP Models
Reza Abbasi (Sharif University of Technology), Mahdieh Soleymani Baghshah (Sharif University of Technology)
CodeExplainability and InterpretabilityRepresentation LearningVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Constructed a novel OoD dataset, ImageNet-AO, specifically designed to evaluate the compositional generalization ability of CLIP models in single-object scenarios, and conducted zero-shot evaluation on multiple CLIP variants using this dataset.
Decomposed Vector-Quantized Variational Autoencoder for Human Grasp Generation
zhao zhe, Huadong Ma (State Key Laboratory of Networking and Switching Technology)
CodeGenerationAuto EncoderPoint Cloud
🎯 What it does: Proposed a decomposition-based vector-quantized variational autoencoder (DVQ-VAE) to generate realistic human grasps that conform to objects.
🎯 What it does: Proposes a decomposition-based test-time optimization method called DecoMotion for precise long-range motion tracking of each pixel in videos.
🎯 What it does: Proposed a decomposition-based discrete neural mapping method (DNMap), which constructs efficient 3D environment representations by learning combinable discrete embeddings and low-resolution continuous embeddings in sparse octrees;
🎯 What it does: This paper proposes a multi-modal self-supervised learning framework called DeCUR, which learns cross-modal consistency and modality-specific information by dividing the feature dimension into common and unique parts, and combines cross-modal and single-modal redundancy reduction loss;
🎯 What it does: This paper proposes the Deep Diffusion Image Prior (DDIP) framework and designs an efficient 3D OOD adaptation method, D3IP, addressing the high computational cost and memory consumption of traditional SCD in 3D inverse problems, while achieving cross-slice joint optimization to improve reconstruction consistency and speed.
🎯 What it does: Propose a decentralized probabilistic aggregation clustering algorithm called PAC, extend it to an online probabilistic aggregation module OPA, and construct an end-to-end deep clustering framework DPAC, which can achieve stable online clustering without relying on center updates;
Lahav Lipson (Princeton University), Jia Deng (Princeton University)
CodeDepth EstimationConvolutional Neural NetworkRecurrent Neural NetworkSimultaneous Localization and MappingOptical FlowImageVideo
🎯 What it does: Proposes DPV-SLAM, a monocular visual SLAM system based on DPVO, incorporating approximate loop closure mechanisms and classical image retrieval loop closure to improve accuracy and robustness.
🎯 What it does: Investigated the vulnerability of document tampering localization (DTL) models under adversarial attacks, and proposed a latent manifold-based adversarial training method to enhance their robustness.
denoiSplit: a method for joint microscopy image splitting and unsupervised denoising
Ashesh Ashesh (Fondazione Human Technopole), Florian Jug (Fondazione Human Technopole)
CodeRestorationSegmentationAuto EncoderBiomedical Data
🎯 What it does: Proposes denoiSplit, a variational segmentation encoder-decoder network capable of simultaneously performing semantic image splitting and denoising under unsupervised conditions.