🎯 What it does: This paper proposes a unified anomaly detection framework named HGAD, which models multi-class normal data using a hierarchical Gaussian mixture regularized normalizing flow, and further enhances inter-class discriminability through mutual information maximization and multi-center learning, thereby achieving detection and localization of multi-class anomalies within a single model.
Hierarchical Separable Video Transformer for Snapshot Compressive Imaging
Ping Wang (Zhejiang University), Xin Yuan (Westlake University)
CodeRestorationCompressionTransformerVideo
🎯 What it does: Proposes a HiSViT architecture based on a hierarchical separable video Transformer for video reconstruction in snapshot compressive imaging (SCI).
🎯 What it does: Proposes a Hierarchical Temporal Context Learning (HTCL) framework based on cameras for completing 3D semantic scene completion from sparse RGB frames.
HiFi-Score: Fine-grained Image Description Evaluation with Hierarchical Parsing Graphs
Ziwei Yao (Chinese Academy of Sciences), Xilin Chen (Chinese Academy of Sciences)
CodeSegmentationExplainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed a HiFi-Score image caption evaluation metric based on hierarchical parsed graphs (HPG), which evaluates the consistency and completeness between text and images through fine-grained matching, and assesses linguistic fluency using a Large Language Model (LLM).
High-Quality Mesh Blendshape Generation from Face Videos via Neural Inverse Rendering
Xin Ming (Tsinghua University), Feng Xu (Tsinghua University)
CodeGenerationVideoMesh
🎯 What it does: Automatically generate personalized mesh blendshape models from single or sparse multi-view facial videos using neural inverse rendering.
🎯 What it does: Propose DL-GS, which leverages the different focal length information from the asynchronous dual-camera system of a smartphone to achieve few-shot and high-resolution view synthesis.
HiT-SR: Hierarchical Transformer for Efficient Image Super-Resolution
XIANG ZHANG (ETH ZURICH), Fisher Yu
CodeSuper ResolutionTransformerImageBenchmark
🎯 What it does: Propose HiT-SR, a general strategy to convert conventional SR Transformers into hierarchical Transformers, efficiently aggregating multi-scale features for image super-resolution through hierarchical windows and spatial-channel correlation methods.
How to Train the Teacher Model for Effective Knowledge Distillation
Shayan Mohajer Hamidi (University of Waterloo), Ahmed Hussein Salamah (University of Waterloo)
CodeClassificationKnowledge DistillationImage
🎯 What it does: Proposed and verified that using mean squared error (MSE) loss to train the teacher model can enhance knowledge distillation (KD) effectiveness, replacing the traditional cross-entropy (CE) teacher with an MSE teacher in various KD methods, and observed performance improvements in the student model.
HowToCaption: Prompting LLMs to Transform Video Annotations at Scale
Nina Shvetsova (Goethe University Frankfurt), Hilde Kuehne (University of Oxford)
CodeData SynthesisRetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextAudio
🎯 What it does: We leverage large language models (e.g., Vicuna-13B) to prompt ASR subtitles from a large number of instructional videos, generating video subtitles in a human-written style and predicting timestamps for each subtitle, thereby constructing a dataset of 25M video-text pairs named HowToCaption without requiring manual annotation.
🎯 What it does: Proposed the HPFF method, combining hierarchical local supervision learning with patch feature fusion to enhance local learning performance and reduce GPU memory consumption.
CodeObject DetectionGenerationData SynthesisAnomaly DetectionVision Language ModelDiffusion modelImageMultimodalityBenchmark
🎯 What it does: Constructed a large-scale human abnormal benchmark AbHuman, and proposed HumanRefiner, an end-to-end coarse-to-fine, reversible pose-guided generation process, significantly improving the accuracy of text-to-image models in human generation.
🎯 What it does: Propose a generative motion model called HUMOS based on body shape conditions that can generate natural, physically feasible, and dynamically stable human motions for different body shapes in one go.
🎯 What it does: This paper proposes the HyperSpaceX framework, which simultaneously explores radial and angular features in multiple hyperspheres, and introduces the DistArc loss to enhance performance in image classification and face recognition.
CodeSegmentationTransformerSupervised Fine-TuningImageBiomedical Data
🎯 What it does: Propose I-MedSAM, a medical image segmentation framework that combines the Segment Anything Model (SAM) with implicit neural representations (INR)
🎯 What it does: Propose the IAM-VFI method, which classifies regions with different motion complexity in videos and generates motion complexity maps using MCENet to achieve high-quality video frame interpolation for arbitrary motion.
🎯 What it does: Proposes an unsupervised skeleton action recognition method based on the Isometric Generative Model (IGM), combining generative pre-training with contrastive learning to achieve isometric constraints in the feature space.
🎯 What it does: Proposes a visual collaborative perception framework based on Instance-Level Fusion Transformer (IFTR), which enhances the 3D detection performance of multi-vehicle cameras by leveraging shared instance features.
Image Compression for Machine and Human Vision With Spatial-Frequency Adaptation
Han Li (Shanghai Jiao Tong University), Hongkai Xiong (Shanghai Jiao Tong University)
CodeCompressionConvolutional Neural NetworkImage
🎯 What it does: A lightweight adapter framework called Adapt-ICMH is studied, which transfers a pre-trained human visual image compression model to machine vision tasks while balancing bitrate and task accuracy.
🎯 What it does: Propose an image demoiréing network (RRID) that leverages dual-domain information from RAW and sRGB. It achieves multi-scale stripe removal by inserting a Gated Feedback Module (GFM) and a Frequency Selection Module (FSM) into the U-Net structure, and restores color through RGB Guided ISP (RGISP) that learns device-specific ISP.
Images are Achilles' Heel of Alignment: Exploiting Visual Vulnerabilities for Jailbreaking Multimodal Large Language Models
Yifan Li (Renmin University of China), Ji-Rong Wen (Renmin University of China)
CodeSafty and PrivacyAdversarial AttackPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Conduct systematic experiments on the harmless alignment of multimodal large language models (MLLMs), demonstrating that image inputs act as backdoors for alignment, and proposing a three-phase automatic attack method HADES to hide and amplify harmful information in images for exploitation.
🎯 What it does: Propose a debiased estimator based on Total Variation (TV) minimization, and use it to construct pixel-level confidence intervals to achieve confidence quantification in undersampled magnetic resonance imaging (MRI).
IMMA: Immunizing text-to-image Models against Malicious Adaptation
Amber Yijia Zheng (Purdue University), Raymond A. Yeh (Purdue University)
CodeGenerationSafty and PrivacyMeta LearningDiffusion modelImageText
🎯 What it does: Before releasing pre-trained text-to-image diffusion models, immunization training is conducted to make the models difficult to adapt to malicious fine-tuning (e.g., replicating artistic styles, recovering deleted concepts, and personalizing content).
🎯 What it does: Proposed a geometry-based implicit concept removal method called Geom-Erasing to eliminate implicit concepts such as watermarks, QR codes, and text in diffusion models.
🎯 What it does: Fine-tune the source model to make its predictions consistent with an independently trained witness model, generating adversarial perturbations with higher transferability.
🎯 What it does: Interpolate potential diffusion models (LDM) during the inference phase, proposing the Contextualized Vendi Score Guidance (c-VSG) mechanism, which utilizes a memory pool and a small number of real samples to guide the generation process, making generated images in different regions more geographically diverse.
🎯 What it does: Introduce Gromov-Wasserstein regularization into existing hyperbolic neural networks (HNNs), enabling the model to better preserve the original geometric structure when mapping Euclidean features to hyperbolic space.
🎯 What it does: Proposed a trainable concept intervention relocalization module to enhance the intervention effectiveness of concept bottleneck models (CBM/CEM/IntCEM) during testing.
🎯 What it does: This paper enhances knowledge distillation by regularizing the feature direction and norm in the embedding layer of the student network, and proposes the Dino-loss loss function;
🎯 What it does: This paper proposes a pixel-level and block-level multi-view consistency loss in the feature space by combining the pre-trained feature priors extracted from multi-view images with Neural Surface Reconstruction (NSR). It systematically evaluates 13 models across seven types of pre-training tasks (MIM, IC, SS, MDE, SM, IM, MVS) and constructs two improved NSR variants, MVS-NeuS and Match-NeuS.
🎯 What it does: Proposes a two-stage text-guided object inpainting method called CAT-Diffusion, which first pre-fills the semantic features of the target object in a multi-modal feature space, and then guides the diffusion model to generate high-fidelity objects through a reference adapter layer.
Improving Virtual Try-On with Garment-focused Diffusion Models
Siqi Wan (University of Science and Technology of China), Tao Mei (HiDream.ai Inc)
CodeImage TranslationImage HarmonizationGenerationVision Language ModelDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: Proposes GarDiff, a garment-focused diffusion model for virtual try-on, capable of generating high-quality, detail-rich wearable effect images.
Improving Vision and Language Concepts Understanding with Multimodal Counterfactual Samples
Chengen Lai (Xidian University), Guangneng Hu (Xidian University)
CodeData SynthesisTransformerVision Language ModelDiffusion modelGenerative Adversarial NetworkContrastive LearningMultimodality
🎯 What it does: By automatically generating multimodal adversarial examples (text and image) and incorporating them into contrastive learning training, the performance of vision-language models in concept understanding and compositional reasoning is improved.
Improving Zero-shot Generalization of Learned Prompts via Unsupervised Knowledge Distillation
Marco Mistretta (University of Florence), Andrew D. Bagdanov (University of Florence)
CodeKnowledge DistillationRepresentation LearningPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: Propose an unsupervised knowledge distillation prompt learning framework KDPL to improve the zero-shot generalization capability of lightweight vision-language models (VLMs).
🎯 What it does: This paper proposes Inf-DiT, a diffusion transformer-based infinite-resolution image upscaling model that achieves efficient upscaling for images of arbitrary sizes using unidirectional block attention (UniBA).
🎯 What it does: Propose an SDF-NeRF framework that utilizes mutual information to constrain surface normals under sparse viewpoints, significantly improving the quality of 3D surface reconstruction.
🎯 What it does: This paper constructs the first large-scale insect identification benchmark dataset, AMI, containing 2.5 million images collected by humans and 2,893 images captured by automated insect traps, with a total of 5,364 moth species and 52,948 individuals annotated. The dataset aims to evaluate the performance of fine-grained insect recognition and domain transfer in field deployment.
Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation
Arpit Garg (University of Adelaide), Gustavo Carneiro (University of Surrey)
CodeClassificationImage
🎯 What it does: Propose a noise rate estimation method based on a graphical model to improve the sample selection curriculum in instance-dependent noise label learning, and seamlessly integrate with existing state-of-the-art (SOTA) noise label learning (LNL) methods.
🎯 What it does: Utilize the DDIM reverse process to convert a single style reference image into noise, and sample from this noise to generate new stylized images, while learning generalizable style tokens through prompt refinement.
InstructIR: High-Quality Image Restoration Following Human Instructions
Marcos V. Conde (University of Würzburg), Radu Timofte (University of Würzburg)
CodeRestorationConvolutional Neural NetworkLarge Language ModelPrompt EngineeringVision Language ModelImageText
🎯 What it does: Proposed a universal image restoration model called InstructIR based on human natural language instructions, capable of restoring images in various tasks such as noise reduction, de-raining, de-fogging, deblurring, and low-light enhancement.
Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection
Xinhao Luo (Institute of Automation, Chinese Academy of Sciences), Guoqi Li (Institute of Automation, Chinese Academy of Sciences)
CodeObject DetectionConvolutional Neural NetworkSpiking Neural NetworkImageTime Series
🎯 What it does: Propose a spiking neural network framework named SpikeYOLO for high-performance, low-energy consumption object detection, and design an I-LIF neuron with integer-value training and spike-driven inference; implemented and verified on the COCO 2017 static dataset and Gen1 neuromorphic event stream dataset.
Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition
Yisong Wang (Peking University), Junsong Yuan (University at Buffalo)
CodeRecognitionConvolutional Neural NetworkRecurrent Neural NetworkGraph Neural NetworkTransformerVision Language ModelVideo
🎯 What it does: This paper proposes an interactive center spatiotemporal context reasoning framework for human-object interaction (HOI) recognition in multi-person videos.
Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection
Yongwei Nie (South China University Of Technology), Hongmin Cai (South China University Of Technology)
CodeAnomaly DetectionAuto EncoderVideo
🎯 What it does: An unsupervised video anomaly detection framework was studied, achieving label-free learning by alternately training a one-class classification (OCC) model and a weakly supervised (WS) model to generate pseudo-labels.
InternVideo2: Scaling Foundation Models for Multimodal Video Understanding
Yi Wang (OpenGVLab, Shanghai AI Laboratory), Limin Wang (OpenGVLab, Shanghai AI Laboratory)
CodeRecognitionRepresentation LearningTransformerLarge Language ModelContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: This paper proposes InternVideo2, a large-scale video foundation model constructed through three-stage progressive training (occlusion-free reconstruction, cross-modal contrastive learning, next token prediction), achieving state-of-the-art (SOTA) performance on multiple video tasks.
Akshay Kulkarni (UC San Diego), Tsui-Wei Weng (UC San Diego)
CodeExplainability and InterpretabilityAdversarial AttackImageBenchmark
🎯 What it does: Proposes a training-agnostic test-time adversarial defense method that leverages neuron interpretability information to rank the importance of internal neurons, retaining only those related to the predicted class during forward propagation;
CodeComputational EfficiencyRepresentation LearningTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Propose a routing function used in vision-language parameter-efficient fine-tuning (VL-PEFT) to better align visual and language features in low-rank bottlenecks;
🎯 What it does: This paper jointly optimizes the camera pose and scene reconstruction in NeRF, proposing to use invertible neural networks (INNs) to over-parameterize the camera pose, thereby improving optimization convergence and pose accuracy.
Yidan Zhang (Beijing Normal University), Baining Guo (Tsinghua University)
CodeRetrievalTransformerImage
🎯 What it does: IRGen treats the image retrieval task as a generation task, using a sequence-to-sequence model to directly output the image identifier of the nearest neighbor of the query image, achieving end-to-end differentiable search.
Is Retain Set All You Need in Machine Unlearning? Restoring Performance of Unlearned Models with Out-Of-Distribution Images
Jacopo Bonato (Leonardo Labs), Luigi Sabetta (Leonardo Labs)
CodeSafty and PrivacyKnowledge DistillationImage
🎯 What it does: Propose the SCAR algorithm that achieves machine model forgetting (unlearning) without requiring a retention set, by migrating the feature vectors of samples to be forgotten to the distribution of the nearest incorrect class to remove information, and using knowledge distillation techniques to maintain model performance on OOD data.
Is user feedback always informative? Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data
Junha Song (KAIST), Jaegul Choo (KAIST)
CodeDomain AdaptationConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: Under the source-free semi-supervised domain adaptation (SemiSDA) framework where source data is inaccessible, this paper investigates the potential negative bias in user feedback (Negatively Biased Feedback, NBF) and proposes a pluggable retrieval-based latent defending (RLD) method to mitigate the adverse effects of NBF on model adaptation performance.
Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models
Xiao Liu (Wuhan University), Jiaxu Miao (Sun Yat-sen University)
CodeGenerationDiffusion modelImage
🎯 What it does: Propose an iterative ensemble training with anti-gradient control (IET-AGC) framework, which weakens memorization in diffusion models by sharding data to train multiple models, periodically aggregating parameters, and dynamically removing low-loss samples during training.
🎯 What it does: Propose the ITTakesTwo framework, which leverages different LiDAR representations from the same scene for consistency learning and contrastive learning, thereby significantly improving the performance of semi-supervised semantic segmentation.
🎯 What it does: This paper proposes and implements a LiDAR-based attention joint detection and tracking framework called JDT3D, and verifies its performance on nuScenes.
Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography
Kailai Zhou (Nanjing University), Xun Cao (Nanjing University)
CodeRestorationConvolutional Neural NetworkTransformerMixture of ExpertsImageMultimodality
🎯 What it does: This paper proposes a Joint RGB-Spectral Decomposition Model Guided Image Enhancement Framework (JDM-HDRNet), which decomposes low-resolution multispectral images (Lr-MSI) and RGB images into shadow, reflection, and material priors jointly, and introduces these priors into HDRNet for tone mapping, local brightness adaptation, and semantic mesh expert learning;
KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter
Yifan Zhan (University of Tokyo), Yinqiang Zheng (Kyoto University)
CodeGenerationNeural Radiance FieldVideo
🎯 What it does: Propose KFD-NeRF, a dynamic NeRF integrated with a Kalman filter, achieving high-quality time-consistent rendering through smooth tri-plane encoding and shallow MLP;
CodeClassificationRetrievalKnowledge DistillationRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodalityBiomedical Data
🎯 What it does: This paper investigates leveraging structured pathological knowledge to enhance visual-language pre-training, aiming to improve performance in retrieval, zero-shot classification, and whole-slide tumor subtyping tasks in computational pathology.
🎯 What it does: Proposes a label-guided cross-modal knowledge distillation method that combines LiDAR teacher models and label features to enhance camera-based 3D object detection performance, while preserving semantic information of the student model through feature partitioning.
LaMI-DETR: Open-Vocabulary Detection with Language Model Instruction
Penghui Du (Beihang University), Si Liu (Baidu)
CodeObject DetectionFederated LearningRepresentation LearningConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose a LaMI-DETR framework based on DETR, which generates visual descriptions using language model instructions (GPT-3.5 + T5), constructs visual concepts and cross-category relationships, thereby improving concept representation and reducing overfitting to base classes.
🎯 What it does: Propose the end-to-end LaneGAP method, which directly detects continuous road paths from camera inputs and converts paths into complete lane graphs through Path2Graph.
Language-Assisted Skeleton Action Understanding for Skeleton-Based Temporal Action Segmentation
Haoyu Ji (Harbin Institute of Technology), Honghai Liu (Harbin Institute of Technology)
CodeSegmentationGraph Neural NetworkTransformerLarge Language ModelVision-Language-Action ModelContrastive LearningTextGraph
🎯 What it does: Propose a language-assisted skeleton action segmentation method called LaSA, which models the semantic relationships between skeleton joints and actions by leveraging language priors, and enhances action classification and boundary prediction performance through contrastive learning with aligned text descriptions.
🎯 What it does: Proposed a RGB-based category-level object pose estimation framework named LaPose, which models object shape uncertainty through a Laplacian Mixture Model and introduces a scale-agnostic pose representation to achieve precise estimation of 9DoF poses.
🎯 What it does: Automatically learn distribution-aware prompts using only the ID category names to enhance the performance of vision-language models in OOD detection.
🎯 What it does: This paper proposes a Deep Unfolding Network (DUN) combining the prior of Latent Diffusion Models (LDM) for three-dimensional spectral image reconstruction in single-shot compressive spectral imaging (CASSI).
🎯 What it does: Built a large-scale amortized text-to-enhanced 3D generation framework called Latte3D, capable of generating high-quality, textured 3D meshes from a single text prompt within 400ms, and supports fast test-time optimization and stylization.
🎯 What it does: This paper proposes a source-free active domain adaptation (SFADA) framework called Learn from the Learnt (LFTL), which addresses scenarios where source data is unavailable and only a minimal amount of target annotations are available.
Learned HDR Image Compression for Perceptually Optimal Storage and Display
Peibei Cao (City University of Hong Kong), Kede Ma (Xellar Biosystems)
CodeCompressionConvolutional Neural NetworkImage
🎯 What it does: Designed and implemented an end-to-end learning HDR image compression framework called EPIC-HDR, achieving dual optimization for HDR image storage and display, generating LDR images compatible with LDR displays and side information for HDR reconstruction.
🎯 What it does: Propose an unsupervised video anomaly detection method that utilizes the first and last frames of a video as prior knowledge of normality, and combines normal propagation with loss-weighted self-training to achieve pseudo-label learning.
🎯 What it does: Propose a weak semi-supervised hidden object detection framework (WSSCOD), which generates high-quality pseudo labels using box prompts and employs a small amount of pixel-level annotations for supervision. A noise correction loss L_NC is introduced to train the main network against pseudo label noise.
🎯 What it does: To address super-resolution tasks with arbitrary scaling ratios in real-world scenarios, the authors constructed the RealArbiSR dataset, which includes both integer and non-integer scaling ratios, and proposed the Dual-Layer Deformable Implicit Representation (DDIR) model. This model simultaneously learns image-level and pixel-level degradation deformations, enabling arbitrary-scale super-resolution under real-world noise degradation conditions.
Learning Equilibrium Transformation for Gamut Expansion and Color Restoration
Jun Xiao (Hong Kong Polytechnic University), Kin-Man Lam (Hong Kong Polytechnic University)
CodeRestorationConvolutional Neural NetworkImage
🎯 What it does: Proposed a lightweight equipotential transformation network for color gamut expansion and color restoration, without relying on any external metadata, implemented through a root search framework and implicit loop mechanism;
Learning Exhaustive Correlation for Spectral Super-Resolution: Where Spatial-Spectral Attention Meets Linear Dependence
Hongyuan Wang, Youliang Yan (Huawei Noah's Ark Lab)
CodeSuper ResolutionTransformerImage
🎯 What it does: Proposes the Exhaustive Correlation Transformer (ECT) to achieve spectral super-resolution restoration from RGB images to hyperspectral images.
🎯 What it does: Propose VectorFormer, a 3D object detection framework based on multi-camera images, which utilizes high-resolution vector representation (Vector Query) combined with low-resolution BEV to achieve fine-grained modeling of 3D space while maintaining efficient inference.
Learning Local Pattern Modularization for Point Cloud Reconstruction from Unseen Classes
Chao Chen (Tsinghua University), Zhizhong Han (Wayne State University)
CodeGenerationAuto EncoderPoint Cloud
🎯 What it does: This paper proposes to achieve point cloud reconstruction for unseen categories by learning local pattern modularization in the object-centered coordinate system, resulting in high-fidelity reconstruction outcomes.
Learning Multimodal Latent Generative Models with Energy-Based Prior
Shiyu Yuan (Stevens Institute of Technology), Tian Han (Stevens Institute of Technology)
CodeGenerationMixture of ExpertsAuto EncoderMultimodalityStochastic Differential Equation
🎯 What it does: Propose a multi-modal latent generative model based on energy-based models (EBM), leveraging EBM prior to enhance latent space expressiveness, achieving semantically consistent generation and cross-modal generation.
🎯 What it does: Pre-train images using metadata such as time and location from satellite images to obtain more semantically expressive visual features.
🎯 What it does: This paper proposes an efficient method to construct the model 'soup'—Memory-Efficient Hyperplane Learned Soup (MEHL-Soup)—on a single GPU, addressing the memory and computational bottlenecks of traditional Learned-Soup approaches.
🎯 What it does: This paper proposes a method called Semantic Latent Directions (SLD) to construct a semantic latent space in stochastic human motion prediction (SHMP), enabling more accurate, controllable, and diverse future motion prediction.
🎯 What it does: Proposes a multi-stage contrastive learning framework (MCL), which gradually uncovers suppressed features through feature-aware negative sampling, and retains all learned features by cross-stage concatenation in the final stage, addressing feature suppression issues in both single-modal and multi-modal contrastive learning.
Learning to Complement and to Defer to Multiple Users
Zheng Zhang (University of Surrey), Gustavo Carneiro (University of Surrey)
CodeClassificationImageBiomedical Data
🎯 What it does: Proposes a unified human-machine collaborative classification framework named LECODU, integrating learning compensation, learning delay, and optimal user number decision-making, and achieving dual optimization of accuracy and collaboration cost through multi-noise label training.
🎯 What it does: Designed a unified anomaly detection framework called OneNIP based on a single normal image prompt, utilizing self-attention reconstruction and cross-attention recovery, combined with supervised refinement for pixel-level anomaly segmentation.
🎯 What it does: This paper proposes the NGCN model based on neighbor information and the cross-view consistency strategy CVCS, improving pseudo-label generation and noise filtering in the general category discovery task.
Learning to Enhance Aperture Phasor Field for Non-Line-of-Sight Imaging
In Cho (Yonsei University), Seon Joo Kim (Yonsei University)
CodeRestorationData SynthesisSuper ResolutionConvolutional Neural NetworkAuto EncoderImagePhysics Related
🎯 What it does: Propose a light phase field enhanced network named LEAP, which recovers complete and clean measurement data from noisy partial observations in sparse sampling and small aperture non-line-of-sight (NLOS) imaging, and further performs scene reconstruction.
🎯 What it does: This paper proposes a bidirectional recursive framework for reconstructing high-speed dynamic scenes from spike streams under low-light conditions, while simultaneously constructing synthetic and real datasets tailored for low-light high-frame-rate scenarios.
🎯 What it does: Propose a robust framework for audio-visual question answering (AVQA) that recovers missing modal features through an RMM generator when input modalities are missing, and further enhances audio-visual features using an AVR diffusion model to achieve accurate answers.
Learning Video Context as Interleaved Multimodal Sequences
Kevin Qinghong Lin (National University of Singapore), Mike Zheng Shou (National University of Singapore)
CodeClassificationRetrievalTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Introduces MovieSeq, a general video understanding framework that embeds narrative videos into interleaved multimodal sequences and uses large language models for instruction tuning.
Learning with Counterfactual Explanations for Radiology Report Generation
Mingjie Li (Stanford University), Xiaojun Chang (MBZUAI)
CodeGenerationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringContrastive LearningImageTextBiomedical Data
🎯 What it does: This paper proposes a pulmonary radiology report generation framework based on contrastive learning and counterfactual explanations (CoFE), which leverages counterfactual images and learnable prompts to enhance the model's focus on abnormal features, generating more accurate and complete medical reports.
🎯 What it does: Proposes a LUT method that incorporates supervision for unmasked tokens in Masked Image Modeling (MIM) training to enhance the representation learning of vision Transformers.
🎯 What it does: Propose a length-aware 3D human motion synthesis framework called LADiff, which can generate corresponding motion sequences based on text descriptions while controlling the target duration.
🎯 What it does: This paper studies how to leverage LiDAR (LiDAR) point clouds during the training phase to enhance the 3D object detection performance of systems that use only imaging radar (3+1D Radar), proposing two methods: Multi-Stage Sparse Training (MSTM) and Cross-Modal Knowledge Distillation (KD).
CodeAutonomous DrivingTransformerSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes a novel online vectorization method for high-definition maps called MapQR, based on the DETR structure. It employs scatter-and-gather queries and position embeddings to simultaneously model content and location information, and improves the BEV encoder (GKT-h) to enhance performance.
🎯 What it does: Proposed a novel data availability attack method called IRP to prevent models from learning unauthorized data in supervised and self-supervised learning scenarios.
CodeAnomaly DetectionTransformerVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes a method for detecting synthetic images by leveraging representations from intermediate Transformer blocks of the CLIP image encoder, combined with a lightweight network and a trainable importance estimator.
Leveraging scale- and orientation-covariant features for planar motion estimation
Marcus Valtonen Örnhag (Ericsson Research), Alberto Jaenal (Ericsson Research)
CodePose EstimationAutonomous DrivingImage
🎯 What it does: This paper derives linear constraints for planar motion from scale and orientation covariant features (e.g., SIFT), proposes a minimization solver that requires only a single SIFT correspondence point, and integrates it into a robust estimation framework;