🎯 What it does: This paper proposes a text-aware masked image modeling (TMIM) pre-training framework based on weakly supervised text detection labels to enhance the performance of scene text removal (STR) models.
CodeClassificationRecognitionData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Proposed a multi-modal large language model specifically for remote sensing image understanding, named LHRS-Bot, and constructed a large-scale remote sensing image-text pair dataset LHRS-Align, an instruction-based dataset LHRS-Instruct, and an evaluation benchmark LHRS-Bench.
🎯 What it does: This paper proposes a 3D detection framework using only LiDAR, which maintains high-precision detection under both normal and adverse weather conditions by incorporating 4D radar prompt learning and four-level cross/inner-modal knowledge distillation during the training phase.
LightenDiffusion: Unsupervised Low-Light Image Enhancement with Latent-Retinex Diffusion Models
Hai Jiang (Sichuan University), Shuaicheng Liu (University of Electronic Science and Technology of China)
CodeRestorationDiffusion modelImage
🎯 What it does: Propose an unsupervised low-light image enhancement framework called LightenDiffusion, combining Retinex theory with diffusion models, utilizing a content transfer decomposition network in the latent space to achieve purer reflection and illumination separation, and improving reconstruction quality through self-constrained consistency loss.
🎯 What it does: Propose an unsupervised linear controllable GAN (LC-GAN), achieving fine-grained semantic control over image generation by decomposing noise vectors into geometric and appearance codes, along with spectral regularization, without relying on pre-trained classifiers or supervisory signals.
LISO: Lidar-only Self-Supervised 3D Object Detection
Stefan Andreas Baur (Mercedes-Benz), Andreas Geiger (Mercedes-Benz)
CodeObject DetectionAutonomous DrivingPoint Cloud
🎯 What it does: Proposes LISO, a self-supervised 3D object detection framework that uses only LiDAR point cloud sequences, leveraging self-supervised scene flow to generate high-precision pseudo labels, and enhancing detection performance through trajectory-regularized self-training iterations;
LITA: Language Instructed Temporal-Localization Assistant
De-An Huang (Nvidia), Jan Kautz (Nvidia)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoText
🎯 What it does: Proposed a specialized temporal localization assistant LITA for video large language models, capable of answering 'when'-type questions and achieving precise time interval localization.
LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models
Yanwei Li (CUHK), Jiaya Jia (CUHK)
CodeComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: Propose LLaMA-VID, which effectively represents video frames through dual tokens (context token and content token), addressing the issue of excessive tokens in Vision-Language Models (VLMs) for long videos.
LMT-GP: Combined Latent Mean-Teacher and Gaussian Process for Semi-supervised Low-light Image Enhancement
Ye Yu (Hefei University of Technology), Zhen Kan (University of Science and Technology of China)
CodeRestorationConvolutional Neural NetworkImage
🎯 What it does: Propose a semi-supervised low-light image enhancement method named LMT-GP, combining the latent mean-teacher framework with Gaussian Process, leveraging unlabelled low-light images to enhance performance and improve generalization.
LNL+K: Enhancing Learning with Noisy Labels Through Noise Source Knowledge Integration
Siqi Wang (Boston University), Bryan Plummer (Boston University)
CodeClassificationImageBiomedical Data
🎯 What it does: This paper proposes incorporating noise source knowledge (LNL+K) into the learning with noisy labels (LNL) task, enhancing the model's ability to identify and train on clean samples through cross-category comparisons of prior information about noise source categories.
Long-CLIP: Unlocking the Long-Text Capability of CLIP
Beichen Zhang (Shanghai AI Laboratory), Jiaqi Wang (Shanghai AI Laboratory)
CodeClassificationRetrievalComputational EfficiencyRepresentation LearningData-Centric LearningTransformerSupervised Fine-TuningVision Language ModelContrastive LearningMultimodality
🎯 What it does: This paper proposes Long-CLIP, a multimodal model that can directly replace CLIP, capable of processing long texts while maintaining or even enhancing zero-shot capabilities.
🎯 What it does: To address the long-tail action distribution in procedural videos, this paper proposes the Group-based Temporal Log Adjustment (G-TLA) framework to improve the segmentation accuracy of tail actions.
CodeRecognitionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: Propose LongVLM, a VideoLLM that achieves efficient fine-grained long video understanding by segmenting long videos, performing hierarchical token merging on each segment, and incorporating global semantics.
🎯 What it does: Proposed and implemented an eye movement scan path prediction model called ART for incremental object reference tasks, and collected a large-scale human eye movement dataset named RefCOCO-Gaze;
🎯 What it does: Propose a fully online framework named BUSCA, which can detect and track targets missed by detectors in any tracking-by-detection (TbD) system, especially maintaining trajectories when targets are occluded.
🎯 What it does: This paper investigates the robustness deficiency of modern deep networks when facing small real image translations, and proposes a new method for achieving robust inference through cropping selection (RICS).
m&m’s: A Benchmark to Evaluate Tool-Use for multi-step multi-modal Tasks
Zixian Ma (University of Washington), Ranjay Krishna (Allen Institute of Artificial Intelligence)
CodeLarge Language ModelAgentic AIPrompt EngineeringMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposes m&m's, a multimodal multi-step tool usage benchmark containing 4K+ real-world tasks and 33 tools, providing a human-validated executable subset.
🎯 What it does: Proposes Masked Conditional Diffusion (MacDiff), a self-supervised skeletal model framework that integrates a semantic encoder with a diffusion decoder.
🎯 What it does: Propose the MagicEraser framework, which completes the object erasure task based on diffusion models, divided into two stages: content initialization (using traditional inpainting pre-filling) and controlled generation (prompt tuning + semantic-aware attention reassembly).
🎯 What it does: Propose the Continual Action Quality Assessment (CAQA) framework, achieving continual learning without forgetting for new and old data through Manifold-Aligned Graph Regularization (MAGR);
Make a Strong Teacher with Label Assistance: A Novel Knowledge Distillation Approach for Semantic Segmentation
Shoumeng Qiu (Fudan University), Jian Pu (Fudan University)
CodeSegmentationKnowledge DistillationImage
🎯 What it does: Introduce noisy labels as auxiliary information for the teacher model in semantic segmentation knowledge distillation, construct a lightweight teacher model, and enhance the student model's performance through distillation.
🎯 What it does: Propose Mamba-ND, a lightweight architecture that extends the 1D state space model Mamba to arbitrary multidimensional data by alternating sequence orders between layers;
MANIKIN: Biomechanically Accurate Neural Inverse Kinematics for Human Motion Estimation
Jiaxi Jiang (ETH Zürich), Christian Holz (ETH Zürich)
CodePose EstimationMesh
🎯 What it does: Proposed MANIKIN, a neural inverse kinematics model based on biomechanical constraints, which infers full-body pose from sparse head and hand poses;
MarvelOVD: Marrying Object Recognition and Vision-Language Models for Robust Open-Vocabulary Object Detection
Kuo Wang (Sun Yat Sen University), Guanbin Li (Sun Yat Sen University)
CodeRecognitionObject DetectionConvolutional Neural NetworkTransformerPrompt EngineeringVision Language ModelImageText
🎯 What it does: This paper proposes the MarvelOVD framework, which leverages collaboration between an object detector and a vision-language model (VLM) to generate high-quality pseudo labels, and enhances open-vocabulary object detection (OVD) performance through online pseudo label mining and adaptive proposal weighting during training.
🎯 What it does: This paper proposes an unsupervised 3D human pose estimation framework based on masks, which achieves direct prediction from a single image to 3D keypoints by using foreground masks as supervision signals.
🎯 What it does: Designed an angle-aware pre-training framework MA3E based on Masked Autoencoder, which learns rotation-invariant features on remote sensing images using rotated cropping and OT loss.
🎯 What it does: This paper proposes a self-supervised point cloud sequence representation learning framework called M2PSC based on Masked Autoencoder, integrating three pre-training tasks: motion trajectory prediction, semantic contrast, and appearance reconstruction.
MasterWeaver: Taming Editability and Face Identity for Personalized Text-to-Image Generation
Yuxiang Wei (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
CodeGenerationTransformerDiffusion modelImageText
🎯 What it does: Proposed a no-tuning method called MasterWeaver for generating personalized images with high identity fidelity and flexible editability.
MaxMI: A Maximal Mutual Information Criterion for Manipulation Concept Discovery
Pei Zhou (University of Hong Kong), Yanchao Yang (University of Hong Kong)
CodeRobotic IntelligenceTransformerSequential
🎯 What it does: This paper proposes a self-supervised framework for discovering key states (manipulation concepts), which identifies important physical states in demonstration trajectories using the MaxMI criterion and employs these states to train concept-guided manipulation strategies.
MedRAT: Unpaired Medical Report Generation via Auxiliary Tasks
Elad Hirsch (Technion Israel Institute of Technology), Ayellet Tal (Technion Israel Institute of Technology)
CodeGenerationTransformerAuto EncoderContrastive LearningImageTextBiomedical Data
🎯 What it does: Proposed a model called MedRAT that generates medical reports from unpaired image and report data, achieving full report generation from X-ray images through autoencoding and multimodal alignment.
Meerkat: Audio-Visual Large Language Model for Grounding in Space and Time
Sanjoy Chowdhury (University of Maryland, College Park), Dinesh Manocha (University of Maryland, College Park)
CodeTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageMultimodalityBenchmarkAudio
🎯 What it does: Developed Meerkat, a multimodal large language model capable of achieving fine-grained audio-visual alignment in both spatial and temporal dimensions, and proposed the AVFIT dataset along with the MeerkatBench unified evaluation framework.
🎯 What it does: Designed and implemented the MAD (Merge-Attend-Diffuse) operator to achieve multi-view interaction in the attention layer of pre-trained diffusion models, thereby generating panoramic images with both visual and semantic consistency.
🎯 What it does: End-to-end learning of high-quality triangular human avatars from multi-view videos, resulting in editable and re-lightable explicit meshes and implicit material fields.
CodeSegmentationVision Language ModelDiffusion modelImageMesh
🎯 What it does: Propose the MeshSegmenter framework to achieve zero-shot grid semantic segmentation based on texture synthesis, by migrating 2D segmentation models to 3D grids and combining multi-view voting to complete unsupervised semantic segmentation.
Meta-Prompting for Automating Zero-shot Visual Recognition with LLMs
Muhammad Jehanzeb Mirza (TU Graz), Horst Possegger (TU Graz)
CodeClassificationRecognitionLarge Language ModelPrompt EngineeringVision Language ModelImageText
🎯 What it does: For zero-shot visual recognition, the MPVR framework is proposed, which automatically generates category-specific VLM prompts through meta-prompts, constructing a diverse set of prompts to achieve zero-shot classification.
Mew: Multiplexed Immunofluorescence Image Analysis through an Efficient Multiplex Network
Sukwon Yun (University of North Carolina at Chapel Hill), Tianlong Chen (University of North Carolina at Chapel Hill)
CodeClassificationGraph Neural NetworkImageBiomedical Data
🎯 What it does: Construct a dual-layer multi-network (Voronoi geometry layer and cell type homogeneity layer), and use an expandable GNN combined with an attention mechanism to perform patient-level multi-fluorescence image phenotype prediction.
🎯 What it does: Propose the MICDrop framework, which enhances unsupervised domain adaptation semantic segmentation performance by adopting complementary dropout and cross-modal feature fusion on RGB and depth features.
🎯 What it does: Propose milliFlow, which uses deep learning to estimate scene flow from millimeter-wave radar point clouds, providing fine-grained motion information for human motion perception and enhancing downstream tasks such as activity recognition, human parsing, and part tracking through this scene flow.
🎯 What it does: Proposes the DIKI framework on visual-language models (e.g., CLIP) to address domain-class incremental learning (DCIL), leveraging residual knowledge integration to preserve pre-trained knowledge and significantly reduce forward forgetting.
Mismatch Quest: Visual and Textual Feedback for Image-Text Misalignment
Brian Gordon (Tel Aviv University), Idan Szpektor (Google Research)
CodeRetrievalAnomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Developed a multi-task model capable of automatically providing textual explanations and visual bounding box annotations when images and text are mismatched, and constructed a large-scale automatically generated training set (TV-Feedback) and a human-annotated evaluation set (SeeTRUE-Feedback).
🎯 What it does: This study addresses the problem of missing modalities in unpaired multimodal data by proposing a parameter-efficient fine-tuning and read-only prompt learning framework based on pre-trained unimodal encoders. The framework can predict missing modality embeddings during inference and fuse them with existing modalities to enhance multimodal classification performance in scenarios with missing modalities.
🎯 What it does: Propose a background class separation framework that reduces catastrophic forgetting caused by background drift in class-incremental semantic segmentation through selective pseudo-labels, adaptive feature distillation, label-guided output distillation, and orthogonal objectives.
ML-SemReg: Boosting Point Cloud Registration with Multi-level Semantic Consistency
Shaocheng Yan (Wuhan University), Jiayuan Li (Wuhan University)
CodePose EstimationAutonomous DrivingPoint Cloud
🎯 What it does: Propose the ML-SemReg framework, which enhances the matching quality of point cloud registration through multi-layer semantic consistency
🎯 What it does: Propose a real-time multi-view single-hand 3D reconstruction method called MLPHand, utilizing a lightweight Skeleton2Mesh model and multi-view geometric feature fusion to achieve high frame rate inference.
MM-SafetyBench: A Benchmark for Safety Evaluation of Multimodal Large Language Models
Xin Liu (East China Normal University), Yu Qiao (Shanghai AI Laboratory)
CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This study proposes MM-SafetyBench, a benchmark for evaluating the safety of multimodal large language models (LLMs), constructing 5,040 text-image pairs across 13 categories of harmful scenarios, and demonstrating the vulnerability of multimodal LLMs through visual prompt attacks.
MMBENCH: Is Your Multi-Modal Model an All-around Player?
Yuan Liu (Shanghai AI Laboratory), Dahua Lin (Shanghai AI Laboratory)
CodeLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Designed and constructed MMBench, a bilingual multimodal model benchmark consisting of over 3,000 multiple-choice questions, to systematically evaluate the perception and reasoning capabilities of vision-language models.
MoAI: Mixture of All Intelligence for Large Language and Vision Models
Byung-Kwan Lee (Korea Advanced Institute of Science and Technology), Yong Man Ro (Korea Advanced Institute of Science and Technology)
CodeCompressionTransformerLarge Language ModelMixture of ExpertsVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a novel large language and vision model called MoAI, which utilizes auxiliary visual information generated by external computer vision models (segmentation, detection, scene graph, OCR). After being compressed by MoAI-Compressor, the auxiliary visual information is fused with visual features and language features through MoAI-Mixer (based on Mixture of Experts) to enhance performance on zero-shot vision-language tasks.
🎯 What it does: Propose ModTr, which converts IR input into RGB-like images through a residual translation network, enabling pre-trained RGB object detectors to directly process new modalities while retaining their original weights.
🎯 What it does: Propose the Model Breadcrumbs method, which merges multiple fine-tuned models based on the same base model through sparse masks to construct a multi-task model without requiring further training.
Model Stock: All we need is just a few fine-tuned models
Dong-Hwan Jang (NAVER AI Lab), Dongyoon Han (NAVER AI Lab)
CodeClassificationKnowledge DistillationConvolutional Neural NetworkSupervised Fine-TuningMixture of ExpertsImage
🎯 What it does: Proposed the Model Stock method, which constructs geometric relationships between a few (e.g., two) fine-tuned models and pre-trained models in the weight space, obtaining fusion weights close to the center to achieve high-quality models;
Modeling and Driving Human Body Soundfields through Acoustic Primitives
Chao Huang (University of Rochester), Alexander Richard (Codec Avatars Lab, Meta)
CodeGenerationMultimodalityAudio
🎯 What it does: The study proposes a 3D human sound field rendering framework based on acoustic primitives, which can predict multiple low-order acoustic primitives from human poses and head-mounted microphone audio, and real-time render full spatial audio.
🎯 What it does: This paper studies the use of coarse correlated equilibrium (CCE) from game theory to model competitive behaviors of vehicles in autonomous driving scenarios under generative world models, thereby generating safety-critical events.
MoE-DiffIR: Task-customized Diffusion Priors for Universal Compressed Image Restoration
Yulin Ren (University of Science and Technology of China), Zhibo Chen (University of Science and Technology of China)
CodeRestorationPrompt EngineeringMixture of ExpertsVision Language ModelDiffusion modelImageBenchmark
🎯 What it does: Propose MoE-DiffIR, a general-purpose compressed image restoration framework based on Stable Diffusion, which dynamically customizes priors for different compression tasks through Mixture-of-Experts prompts and activates SD's cross-modal priors using a visual-to-text adapter;
MoEAD: A Parameter-efficient Model for Multi-class Anomaly Detection
Shiyuan Meng (Zhejiang University), Shibo He (Zhejiang University)
CodeAnomaly DetectionTransformerMixture of ExpertsImage
🎯 What it does: Propose a parameter-efficient ViT-style model MoEAD for unsupervised unified multi-class anomaly detection, adopting Mixture of Experts (MoE) technology to share parameters in the Transformer decoder and achieve finer-grained expert selection through SMoE layers, combined with learnable query embeddings, neighborhood mask attention, and auxiliary loss to balance expert load, ultimately achieving detection and localization of anomalous images;
🎯 What it does: Propose MOFA-Video, which utilizes a frozen Stable Video Diffusion model and multi-domain motion field adapters (MOFA-Adapters) to achieve controllable image animation based on sparse control signals (e.g., hand-drawn trajectories, facial key points, etc.);
🎯 What it does: Designed and verified the Momentum Auxiliary Network (MAN) to enhance inter-block information exchange in supervised local learning, thereby improving model accuracy and reducing memory consumption.
🎯 What it does: Propose a unified self-supervised single-frame and multi-frame monocular depth estimation framework called Mono-ViFI, which utilizes video frame interpolation (VFI) to achieve temporal domain data augmentation and designs a VFI-assisted multi-frame feature alignment and fusion module; on this basis, spatial affine transformation enhancement and triple depth consistency loss are added for mutual distillation and scale self-consistency.
MONTAGE: Monitoring Training for Attribution of Generative Diffusion Models
Jonathan Brokman (Fujitsu Research of Europe), Hisashi Kojima (Fujitsu Limited)
CodeGenerationExplainability and InterpretabilityTransformerVision Language ModelDiffusion modelImage
🎯 What it does: Monitor the generation process of custom diffusion models, construct a data attribution table, and train a distance metric learning model to predict the contribution of training samples on unmonitored generated images.
🎯 What it does: This paper proposes a deviation accumulation event representation method based on event cameras, and designs an end-to-end image deblurring network that integrates a Recursive Motion Extraction (RME) module and a Bidirectional Feature Alignment Fusion (FAF) module.
🎯 What it does: Propose the MTaDCS method, which selects confident samples through movement trajectory and local feature density to address the confusion between noisy labels and difficult samples.
🎯 What it does: Propose a multi-teacher knowledge distillation framework called MTKD, which uses a knowledge aggregation network to fuse outputs from multiple teachers and guide a lightweight student network to learn image super-resolution.
MTMamba: Enhancing Multi-Task Dense Scene Understanding by Mamba-Based Decoders
Baijiong Lin (Hong Kong University of Science and Technology), Yingcong Chen (Hong Kong University of Science and Technology)
CodeSegmentationDepth EstimationTransformerImage
🎯 What it does: Propose a Mamba-based multi-task dense scene understanding framework named MTMamba, which includes two core modules: Self-Task Mamba (STM) and Cross-Task Mamba (CTM), aiming to simultaneously accomplish dense prediction tasks such as semantic segmentation, depth estimation, and human parsing.
🎯 What it does: Propose a multi-branch collaborative learning network (MCLN) that simultaneously handles the 3DREC and 3DRES tasks in 3D visual localization;
🎯 What it does: Designed an end-to-end scene graph generation method based on a multi-granularity sparse relationship matrix prediction network, eliminating the relation matching step and enhancing relation prediction by leveraging entity semantics and positional information.
🎯 What it does: Propose Multi-HMR, a single-frame multi-human full-body 3D mesh recovery model based on Vision Transformer, which can simultaneously detect multiple people, regress SMPL-X parameters, predict 3D positions, and optionally utilize camera intrinsics from a single RGB image.
🎯 What it does: Proposes a Multi-Memory Matching (MMM) framework for unsupervised visible-infrared person re-identification, generating high-quality pseudo labels and establishing reliable cross-modal correspondences through cross-modal clustering, multi-memory learning and matching, and soft clustering-level alignment.
🎯 What it does: This paper proposes a multi-RoI human mesh recovery method that simultaneously estimates the global 3D mesh and local cameras using multiple local views, and improves accuracy through camera consistency loss and contrastive loss.
Jiali Cheng (University of Massachusetts Lowell), Hadi Amiri (University of Massachusetts Lowell)
CodeSafty and PrivacyTransformerMultimodality
🎯 What it does: To address the forgetting task in multimodal machine learning models, this paper proposes the MultiDelete method, which can forget the multimodal associations of specified samples without retraining the entire model, while maintaining the original functionality;
Multimodal Label Relevance Ranking via Reinforcement Learning
Taian Guo (Tencent Youtu Lab), Xing Sun (Tencent Youtu Lab)
CodeDomain AdaptationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Propose a reinforcement learning-based multimodal tag relevance ranking method called LR PPO, and construct a new dataset specifically for this task named LRMovieNet; this method can train a base model on the source domain and quickly transfer to the target domain with only a small amount of preference annotations.
🎯 What it does: Propose the Graph Texture Network (GTN), which maps multi-layer convolutional features into a graph structure and achieves adaptive fusion of latent texture attributes through a learnable undirected masked adjacency matrix and residual message passing;
Multiscale Sliced Wasserstein Distances as Perceptual Color Difference Measures
Jiaqi He (City University of Hong Kong), Kede Ma
CodeImage
🎯 What it does: Propose an image color difference evaluation method based on multi-scale sliced Wasserstein distance (MS-SWD), which uses non-local patch matching for potentially mismatched photographic images to assess color differences.
Multistain Pretraining for Slide Representation Learning in Pathology
Guillaume Jaume (Harvard Medical School), Faisal Mahmood (Mass General Brigham)
CodeClassificationRepresentation LearningTransformerContrastive LearningMultimodalityBiomedical Data
🎯 What it does: This paper proposes Madeleine, a multi-modal self-supervised pre-training framework based on multiple staining (e.g., H&E and IHC), designed to learn general representations for whole slide images (WSI).
🎯 What it does: This paper proposes MutDet, a mutual optimization pre-training framework for orientation object detection in remote sensing images, aiming to address the feature mismatch problem caused by differences in feature extraction methods in traditional DETR pre-training.
🎯 What it does: Propose an efficient multi-view sparse image to 3D Gaussian distribution inference model, MVSplat, which achieves 3D Gaussian representation and novel view rendering with a single forward pass.
NavGPT-2: Unleashing Navigational Reasoning Capability for Large Vision-Language Models
Gengze Zhou (University of Adelaide), Qi Wu (University of Adelaide)
CodeAutonomous DrivingExplainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Develop NavGPT-2 by integrating a frozen large vision-language model with topological graph navigation strategies, enabling interpretable navigation reasoning and interaction in visual environments.
Navigation Instruction Generation with BEV Perception and Large Language Models
Sheng Fan (Zhejiang University), Yi Yang (Zhejiang University)
CodeAutonomous DrivingTransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelImageTextMultimodality
🎯 What it does: Developed a navigation instruction generation framework called BEVInstructor, which integrates BEV perception and large language models (LLM) to generate natural language navigation instructions by fusing multi-view images with bird's-eye-view (BEV) features.
Neural graphics texture compression supporting random access
Farzad Farhadzadeh (Qualcomm AI Research), Fatih Porikli (Qualcomm AI Research)
CodeCompressionTransformerAuto EncoderImage
🎯 What it does: Proposes a neural texture compression framework that supports random access, enabling real-time decoding of multi-channel textures at any MIP level and texture coordinates.
🎯 What it does: Propose a NeRF-based closed-loop simulation framework called NeuroNCAP to test the performance of end-to-end autonomous driving models on photorealistic sensor data in safety-critical scenarios; by rendering nuScenes logs and dynamically editing actors, construct Euro NCAP-inspired three categories of collision scenarios (stationary, forward, lateral) and conduct multiple closed-loop evaluations.
🎯 What it does: By performing 1–3 iterations of noise calibration on random noise, combined with the pre-trained diffusion model SDEdit, an enhancement method is achieved that improves video quality while maintaining video content consistency.
🎯 What it does: Propose a non-sample domain incremental learning method called PINA, achieving cross-domain concept integration through a unified classifier, domain-specific alignment module, and patch shuffling selector.
🎯 What it does: Constructed a large-scale non-verbal interaction dataset called NVI, proposed the NVI-DET task, and designed a detection model named NVI-DEHR based on a dual-scale hypergraph.
Norma: A Noise Robust Memory-Augmented Framework for Whole Slide Image Classification
Yu Bai (Beijing University of Posts and Telecommunications), Wendong Wang (Beijing University of Posts and Telecommunications)
CodeClassificationTransformerContrastive LearningImageBiomedical Data
🎯 What it does: Proposes a Whole Slide Image classification framework called Norma, based on serialized and memory-enhanced Vision Transformer, and enhances robustness through cyclic training and noise label detection.
OAT: Object-Level Attention Transformer for Gaze Scanpath Prediction
Yini Fang (Hong Kong University of Science and Technology), Bertram E Shi
CodeTransformerImage
🎯 What it does: This paper proposes an Object-Level Attention Network (OAT) based on Transformer for predicting human eye movement scan paths in visual search tasks, modeling scan paths as a sequence of object fixations rather than pixel-level fixations.
CodeObject DetectionRetrievalVision Language ModelMultimodality
🎯 What it does: Designed an object-aware perturbation for queries in pre-trained vision-language models to enhance the retrieval capability of small objects in images.
🎯 What it does: Propose an energy-based alignment (EBAMA) method for optimizing attention maps in text-guided diffusion models during inference, along with corresponding object-level attribute binding loss and strength regularization terms.
Occluded Gait Recognition with Mixture of Experts: An Action Detection Perspective
Panjian Huang (Beijing Normal University), Yongzhen Huang (Beijing Normal University)
CodeRecognitionConvolutional Neural NetworkMixture of ExpertsContrastive LearningVideo
🎯 What it does: To address gait recognition in occluded environments, we propose a hybrid expert model called GaitMoE based on action detection, and create the Occluded Gait (OccGait) database.
🎯 What it does: Propose the Occlusion-Aware Seamless Segmentation (OASS) task, construct the UnmaskFormer model to achieve multi-task segmentation (semantic, instance, amodal) on panoramic images, and address three major challenges—field-of-view extension, occlusion, and domain difference—via Unmasking Attention (UA) and Amodal-oriented Mix (AoMix).
🎯 What it does: Propose a 3D occupancy prediction framework called Occupancy as Set of Points (OSP) based on point interest points (PoI), which uses sparse points instead of traditional voxel grids for sampling and prediction.
🎯 What it does: This paper constructs a world model called OccWorld based on 3D occupied space, used for simultaneously predicting future 4D occupancy information and the trajectory of autonomous vehicles.
🎯 What it does: Propose the OGNI-DC framework, which achieves sparse depth completion through differentiable optimization-guided neural iteration, using depth gradients as the learning objective and leveraging DDI integration to obtain complete depth maps.