π― What it does: A weakly supervised video anomaly detection framework PI-VAD has been developed, which enhances RGB representation during the training phase using pseudo-modal generation and cross-modal induction with five additional modalities, while inference only uses RGB.
π― What it does: A novel continuous learning classifier named Kolmogorov-Arnold Classifier (KAC) is proposed, which effectively mitigates catastrophic forgetting while maintaining stable feature distribution.
π― What it does: A pluggable Koopman Multimodal Decomposition (KMD) module is proposed for brain tumor segmentation tasks under varying modal missing rates (balanced or unbalanced), capable of decomposing the features of each modality into shared and exclusive information, and constructing inter-modal relationships.
Knowledge Memorization and Rumination for Pre-trained Model-based Class-Incremental Learning
Zijian Gao (National University of Defense Technology), Huaimin Wang (National University of Defense Technology)
CodeClassificationTransformerImage
π― What it does: In replay-free class-incremental learning, researchers utilize pre-trained models and propose the MoAL method, which integrates momentum adapter weight interpolation, knowledge memory, and a reminiscence mechanism to continuously enhance model adaptability and alleviate catastrophic forgetting.
Language Guided Concept Bottleneck Models for Interpretable Continual Learning
Lu Yu (Tianjin University of Technology), Changsheng Xu (University of Science and Technology of China)
CodeExplainability and InterpretabilityLarge Language ModelVision Language ModelImage
π― What it does: This paper proposes an explainable continual learning framework based on a language-guided Concept Bottleneck Model (CBM), aimed at alleviating catastrophic forgetting while enhancing model interpretability.
π― What it does: A multi-dimensional language-guided audio-visual learning framework MLAVL is proposed for quality assessment of long-duration sports videos.
Large-scale Multi-view Tensor Clustering with Implicit Linear Kernels
Jiyuan Liu (National University of Defense Technology), Ke Liang (National University of Defense Technology)
CodeOptimizationMultimodalityBenchmark
π― What it does: A large-scale multi-view tensor clustering method LMTC is proposed, which eliminates the traditional tensor rotation technique and clusters all samples by embedding their similarities into a low-rank tensor using an implicit linear kernel.
π― What it does: A framework for super-resolution re-sampling in latent space, LSRNA, is proposed, which combines region-adaptive noise injection to enhance the generation quality and speed of text-to-high-resolution image conversion.
π― What it does: Utilizing a deformed Transformer to directly regress and iteratively optimize 3D Gaussian spheres for single-view novel view synthesis.
π― What it does: This paper proposes a cross-attention entropy model based on a shared learnable dictionary (DCAE) to enhance the performance of entropy models in learned image compression.
π― What it does: An end-to-end network based on dense matching and geometric constraints is proposed, capable of simultaneously extracting high-quality point correspondences and local affine transformations to generate accurate affine correspondences.
π― What it does: This paper proposes a unified sparse supervised 3D object detection framework that can achieve object detection with only one box label per scene in both indoor and outdoor environments.
Learning Endogenous Attention for Incremental Object Detection
Xiang Song (Xi'an Jiaotong University), Yihong Gong (Xi'an Jiaotong University)
CodeObject DetectionTransformerImage
π― What it does: This paper proposes an incremental object detection method LEA based on an internal attention mechanism, aimed at addressing the issue of incomplete annotations in incremental detection.
Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing
Ruiyi Wang (Shanghai Jiao Tong University), Xiaohong Liu (Shanghai Jiao Tong University)
CodeRestorationGenerationDiffusion modelImage
π― What it does: A mist removal pipeline based on diffusion models is proposed, which first generates high-quality real fog images using HazeGen, and then performs dehazing using DiffDehaze.
Learning Heterogeneous Tissues with Mixture of Experts for Gigapixel Whole Slide Images
Junxian Wu (Southeast University), Youyong Kong (Southeast University)
CodeClassificationSegmentationGraph Neural NetworkTransformerMixture of ExpertsImage
π― What it does: This paper proposes a pluggable Pathology-Aware Mixture of Experts (PAMoE) module to capture tissue heterogeneity and perform survival prediction on whole slide images (WSI).
π― What it does: To address the occlusion problem in drone tracking, this paper proposes the ORTrack framework based on Vision Transformer, utilizing spatial Cox process random occlusion to achieve occlusion-robust feature learning, and generating a more efficient student model through adaptive feature knowledge distillation;
Learning Phase Distortion with Selective State Space Models for Video Turbulence Mitigation
Xingguang Zhang (Purdue University), Stanley H. Chan (Purdue University)
CodeRestorationAuto EncoderVideo
π― What it does: This study investigates the problem of video turbulence attenuation and proposes the network MambaTM for simultaneous estimation and removal of turbulence.
π― What it does: An unsupervised method is proposed to estimate physical parameters from videos of single-segment continuous dynamical systems. An encoder is used to map frames to latent space, and a physical block predicts the next latent state based on known differential equations, avoiding model collapse through KL divergence loss, without the need to reconstruct a decoder.
π― What it does: This paper proposes an unsupervised multi-agent LiDAR 3D object detection framework called DOtA, which utilizes shared pose and shape information to train an initial detector and enhances the quality of pseudo-labels through multi-scale boundary encoding and label-internal contrastive learning, ultimately training the detector.
Learning to Normalize on the SPD Manifold under Bures-Wasserstein Geometry
Rui Wang (Jiangnan University), Xiao-Jun Wu (Jiangnan University)
CodeClassificationRecognitionConvolutional Neural NetworkTime SeriesBiomedical Data
π― What it does: This paper proposes a General Bures-Wasserstein metric (GBW)-based SPD (Symmetric Positive Definite Matrix) batch normalization (GBWBN), and introduces learnable SPD parameters and power transformations to achieve adaptive normalization of batch features.
Learning with Noisy Triplet Correspondence for Composed Image Retrieval
Shuxian Li (Sichuan University), Peng Hu (Sichuan University)
CodeRetrievalVision Language ModelContrastive LearningImage
π― What it does: This paper addresses the issue of noise in triplets caused by manual annotation (NTC) and proposes a robust learning framework named TME, which can achieve high-quality composite image retrieval in noisy environments.
Learning-enabled Polynomial Lyapunov Function Synthesis via High-Accuracy Counterexample-Guided Framework
Hanrui Zhao (East China Normal University), Zhengfeng Yang (East China Normal University)
CodeOptimizationSupervised Fine-TuningTime Series
π― What it does: A learning-based polynomial Lyapunov function synthesis framework, SynNLF, is proposed, which combines polynomial neural network training with SOS verification and accelerates the learning process through high-precision counterexample generation.
π― What it does: Systematically evaluated and compared various parameter-efficient fine-tuning (PEFT) methods for visual Transformers, conducting unified experiments on their performance in low-sample, sufficient-sample, and distribution shift scenarios.
π― What it does: This paper proposes a post-processing method called PRO, which minimizes the softmax score by performing a limited perturbation search on the input to enhance OOD detection.
Lifting the Veil on Visual Information Flow in MLLMs: Unlocking Pathways to Faster Inference
Hao Yin (University of Science and Technology of China), Zilei Wang (University of Science and Technology of China)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelVision Language ModelTextMultimodality
π― What it does: This paper explores the flow mechanism of visual information at different network depths in multimodal large language models (MLLMs) and proposes a hierarchical visual token pruning method, HiMAP, to achieve inference acceleration.
LiMoE: Mixture of LiDAR Representation Learners from Automotive Scenes
Xiang Xu (Nanjing University of Aeronautics and Astronautics), Qingshan Liu (Nanjing University of Posts and Telecommunications)
CodeSegmentationAutonomous DrivingKnowledge DistillationRepresentation LearningMixture of ExpertsContrastive LearningPoint Cloud
π― What it does: This paper proposes the LiMoE framework, which integrates three types of LiDAR representations (range images, sparse voxels, and raw point clouds) through Mixture of Experts (MoE) to achieve pre-training and downstream segmentation tasks.
LION-FS: Fast & Slow Video-Language Thinker as Online Video Assistant
Wei Li (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
CodeRecognitionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
π― What it does: A real-time online assistant LION-FS based on fast and slow paths is proposed, capable of processing high frame rate first-person videos and generating accurate responses.
π― What it does: A LiSu synthetic LiDAR dataset is proposed along with a single-step surface normal estimation method based on graph total variation regularization, which takes into account spatial and temporal consistency.
LiVOS: Light Video Object Segmentation with Gated Linear Matching
Qin Liu (University of North Carolina Chapel Hill), Lijuan Wang (Microsoft)
CodeSegmentationVision Language ModelVideo
π― What it does: A lightweight video object segmentation network LiVOS is proposed, using linear matching instead of traditional softmax matching, achieving a constant size memory state.
LLaVA-ST: A Multimodal Large Language Model for Fine-Grained Spatial-Temporal Understanding
Hongyu Li (Beihang University), Si Liu (Beihang University)
CodeRecognitionObject DetectionSegmentationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodality
π― What it does: LLaVA-ST is proposed, a multimodal large language model capable of simultaneously handling spatiotemporal fine-grained localization and description of videos.
π― What it does: A linear Mamba operator (LMO) is proposed for MRI reconstruction, achieving the mapping from undersampled k-space to high-quality images.
Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation
Byung Hyun Lee (Seoul National University), Se Young Chun (Seoul National University)
CodeGenerationData SynthesisDiffusion modelImage
π― What it does: This paper studies a training-free local concept elimination method called GLoCE, which can accurately remove target concepts appearing in text prompts while keeping the rest of the image intact.
LOGICZSL: Exploring Logic-induced Representation for Compositional Zero-shot Learning
Peng Wu (Shandong University), Wenguan Wang (University of Macau)
CodeClassificationRepresentation LearningTransformerLarge Language ModelVision Language ModelMultimodality
π― What it does: The LOGICZSL framework is proposed, which incorporates relational knowledge generated by large language models into the training of the CZSL model through logical rules.
π― What it does: Proposes the Logits DeConfusion method, which combines Multi-layer Adapter Fusion (MAF) and Inter-Class Deconfusion (ICD) modules to address the inter-class confusion problem in CLIP for few-shot learning.
LoRA Recycle: Unlocking Tuning-Free Few-Shot Adaptability in Visual Foundation Models by Recycling Pre-Tuned LoRAs
Zixuan Hu (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
CodeDomain AdaptationMeta LearningImageBenchmark
π― What it does: The LoRA Recycle framework achieves few-shot adaptation of visual foundation models without fine-tuning by utilizing pre-finetuned LoRA and generating synthetic data through data-free reverse generation, constructing meta-LoRA using meta-learning.
LoRASculpt: Sculpting LoRA for Harmonizing General and Specialized Knowledge in Multimodal Large Language Models
Jian Liang (Wuhan University), Mang Ye (Wuhan University)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextMultimodality
π― What it does: This paper studies and proposes the LoRASculpt framework, which reduces forgetting and enhances performance of multimodal large language models in downstream tasks through sparse sculpting and conflict mitigation regularization of LoRA.
π― What it does: The LoTUS method is proposed, which achieves forgetting of specified training samples by adjusting the probability distribution in the model output space through entropy regulation, avoiding retraining from scratch.
LotusFilter: Fast Diverse Nearest Neighbor Search via a Learned Cutoff Table
Yusuke Matsui (University of Tokyo)
CodeRetrievalComputational EfficiencyText
π― What it does: This paper proposes LotusFilter, a post-processing module based on a pre-constructed cutoff table, designed to quickly generate diverse results after nearest neighbor retrieval.
π― What it does: Utilize diffusion models to generate a low-bias general annotation dataset using only category names, and pre-train a visual backbone network on this dataset.
LP-Diff: Towards Improved Restoration of Real-World Degraded License Plate
Haoyan Gong (Xi'an Jiaotong-Liverpool University), Hongbin Liu (Xi'an Jiaotong-Liverpool University)
CodeRecognitionRestorationDiffusion modelImage
π― What it does: This paper proposes the LP-Diff network for license plate image restoration based on diffusion models, as well as the first real multi-frame degraded license plate dataset MDLP, addressing the reconstruction and recognition issues of severely degraded license plates in real-world scenarios.
LPOSS: Label Propagation Over Patches and Pixels for Open-vocabulary Semantic Segmentation
Vladan StojniΔ (Czech Technical University in Prague), Giorgos Tolias (NAVER LABS Europe)
CodeSegmentationTransformerVision Language ModelContrastive LearningImage
π― What it does: An open-source, training-free vocabulary semantic segmentation method LPOSS and LPOSS+ is proposed, which refines the initial predictions generated by VLMs (such as CLIP) at the patch and pixel levels through label propagation.
LSceneLLM: Enhancing Large 3D Scene Understanding Using Adaptive Visual Preferences
Hongyan Zhi (South China University of Technology), Chuang Gan (UMass Amherst)
CodeTransformerLarge Language ModelVision Language ModelPoint CloudBenchmark
π― What it does: The LSceneLLM framework is proposed, which utilizes the visual preferences of LLM to automatically locate task-relevant areas and amplifies and fuses fine-grained features through a pluggable scene amplification module, thereby enhancing the understanding capability of large scene 3D visual language models.
π― What it does: To address the high-quality view synthesis problem in multi-view scenes under low light, overexposure, and varying exposure conditions, a method is proposed that incorporates view-adaptive color matrix mapping and curve adjustment within the 3D Gaussian Splatting framework.
MAC-Ego3D: Multi-Agent Gaussian Consensus for Real-Time Collaborative Ego-Motion and Photorealistic 3D Reconstruction
Xiaohao Xu (University of Michigan), Xiaonan Huang (University of Michigan)
CodePose EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingPoint Cloud
π― What it does: This paper proposes a real-time multi-agent collaborative localization and 3D reconstruction framework named MAC-Ego3D, which achieves collaborative pose estimation and high-fidelity dense reconstruction through multi-agent Gaussian consensus.
π― What it does: A continuous testing adaptive method named Topological Consistency Adaptation (TCA) is proposed to continuously and stably adapt to the changing target domain without accessing the source data.
Making Old Film Great Again: Degradation-aware State Space Model for Old Film Restoration
Yudong Mao (City University of Hong Kong), Shiqi Wang (City University of Hong Kong)
CodeRestorationOptical FlowVideoBenchmark
π― What it does: A dynamic degradation-aware restoration framework MambaOFR based on Mamba is proposed to address various mixed degradation issues in old films.
Feng Wang (Johns Hopkins University), Cihang Xie (UC Santa Cruz)
CodeClassificationSegmentationTransformerImage
π― What it does: By uniformly inserting registration tokens into the Vision Mamba model and reusing the registration head at the end, high norm abnormal tokens in the feature map are reduced, enhancing classification and segmentation performance.
Mamba4D: Efficient 4D Point Cloud Video Understanding with Disentangled Spatial-Temporal State Space Models
Jiuming Liu (Shanghai Jiao Tong University), Hesheng Wang (Shanghai Jiao Tong University)
CodeRecognitionSegmentationVideoPoint Cloud
π― What it does: A 4D point cloud video understanding backbone called Mamba4D is proposed based on the State Space Model (Mamba), decoupling space and time, and achieving short-term and long-term spatiotemporal associations through Intra-frame Spatial Mamba and Inter-frame Temporal Mamba.
π― What it does: This paper proposes MambaIRv2, an improved state-space model of Mamba, which incorporates ViT-style non-causal modeling and is applied to image super-resolution, denoising, and JPEG compression loss recovery.
π― What it does: This paper questions the necessity of Mamba in visual tasks and proposes a model called MambaOut that removes SSM. It evaluates this model on common visual benchmarks such as ImageNet classification, COCO detection/segmentation, and ADE20K segmentation.
π― What it does: This paper proposes a hybrid Mamba-Transformer visual backbone called MambaVision, which redesigns the state space module of Mamba and combines it with self-attention blocks for more efficient visual feature modeling.
π― What it does: This paper presents the MammAlps dataset, a set of multimodal, multi-view, long-term wildlife behavior monitoring data, which includes video, audio, scene segmentation, and dense hierarchical behavior labels.
π― What it does: A two-stage MANIPTRANS framework is proposed, first using a large-scale MoCap pre-trained hand trajectory imitation model, and then refining the residual module to achieve precise simulation and task execution of dual robotic hands.
π― What it does: A self-supervised pretraining framework named Masked Autoregressive Pretraining (MAP) is proposed for jointly training a hybrid visual backbone network that combines Mamba and Transformer modules.
MarkushGrapher: Joint Visual and Textual Recognition of Markush Structures
Lucas Morin (IBM Research), Peter Staar (IBM Research)
CodeRecognitionGenerationData SynthesisTransformerVision Language ModelImageTextMultimodality
π― What it does: A multi-modal Markush structure recognition framework called MarkushGrapher is proposed, which can automatically generate graphical representations and variable group tables of Markush structures by utilizing image, text, and layout information simultaneously.
Marten: Visual Question Answering with Mask Generation for Multi-modal Document Understanding
Zining Wang (Meituan), Xiaokang Yang (Shanghai Jiao Tong University)
CodeRecognitionSegmentationOptimizationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
π― What it does: This paper proposes a pre-training task based on visual question answering and mask generation called VQAMask, and trains a multimodal large language model named Marten for document-level visual understanding.
π― What it does: This paper proposes MaskGWM, a driving world model based on DiT, which enhances long-term prediction and multi-view generation performance by incorporating a video mask reconstruction task during training.
π― What it does: The MaskSub framework is proposed, which runs the main branch and the sub-branch in parallel. The sub-branch uses a high ratio of masked inputs and is guided by self-distillation-style relaxed loss during training, allowing supervised learning to withstand strong masked augmentation.
π― What it does: MATCHA is proposed, a unified feature model capable of achieving 'matching anything' across geometric, semantic, and temporal matching tasks.
MBQ: Modality-Balanced Quantization for Large Vision-Language Models
Shiyao Li (Tsinghua University), Yu Wang (Tsinghua University)
CodeCompressionOptimizationTransformerVision Language ModelMultimodality
π― What it does: This paper proposes a post-training quantization method for large visual-language models called MBQ (Modality-Balanced Quantization), which improves quantization quality by considering the different sensitivities to errors of the visual and language modalities during the calibration process.
π― What it does: Through optimization during inference, MC 2 seamlessly integrates multiple trained single-concept custom models (such as Textual Inversion, LoRA, DreamBooth) to generate multi-concept images.
π― What it does: A hybrid mesh-Gaussian representation for full-body animated avatars, MeGA, is proposed, capable of high-quality rendering of faces and hair while supporting editing.
π― What it does: To address the 3D point cloud reconstruction problem from a single-view image, the authors propose a new method called MESC-3D, which guides point cloud generation through effective semantic cues.
Meta-Learning Hyperparameters for Parameter Efficient Fine-Tuning
Zichen Tian (Singapore Management University), Qianru Sun (Singapore Management University)
CodeOptimizationMeta LearningImage
π― What it does: This paper studies and proposes an adaptive parameter-efficient fine-tuning method named MetaPEFT, designed for long-tail tasks in remote sensing and natural images, which can automatically learn the insertion positions, layer depths, and scaling factors of PEFT modules.
π― What it does: This paper proposes MetricGrids, which utilizes multi-dimensional metric grids to approximate continuous signals through Taylor expansion, achieving high-order nonlinear feature interpolation.
π― What it does: A global multi-region multi-satellite sea fog detection and prediction dataset, MFogHub, has been constructed, and multi-model benchmark experiments have been conducted on it.
Mind the Gap: Confidence Discrepancy Can Guide Federated Semi-Supervised Learning Across Pseudo-Mismatch
Yijie Liu (Xiamen University), Hanzi Wang (Xiamen University)
CodeFederated LearningContrastive LearningImage
π― What it does: This study investigates the issue of mismatched pseudo-labels caused by data heterogeneity in federated semi-supervised learning and proposes the SAGE method, which enhances model performance through adaptive pseudo-label correction based on confidence differences.
π― What it does: This paper proposes a GWAD framework based on query update similarity (Delta Similarity) for real-time detection of black-box adversarial attacks.
π― What it does: This paper systematically studies and verifies the feasibility of 'hijacking attacks' on text-to-image diffusion models based on IP-Adapter using image prompts, revealing the potential harm of such attacks to the reputation of service providers and user experience.
π― What it does: This paper proposes a data-driven alternating learning (DALE) framework to enhance the stability and accuracy of medical image lesion segmentation in ambiguous regions.
π― What it does: By constructing a unified cross-modal image matching framework MINIMA and utilizing a large data engine to automatically generate multi-modal paired data from RGB images, the 'modal gap' problem in cross-modal matching is addressed;
π― What it does: This paper studies how to dynamically match activation functions for each input signal in implicit neural representations (INR), constructing a network structure that aligns with the characteristics of the signal.
Missing Target-Relevant Information Prediction with World Model for Accurate Zero-Shot Composed Image Retrieval
Yuanmin Tang (Institute of Information Engineering Chinese Academy of Sciences), Qi Wu (University of Adelaide)
CodeRetrievalContrastive LearningWorld ModelImage
π― What it does: A prediction-based image-to-word mapping network called PrediCIR is proposed, which is based on a world model to predict missing target visual content and map it to pseudo-words in zero-shot synthesized image retrieval.
Mitigating Object Hallucinations in Large Vision-Language Models with Assembly of Global and Local Attention
Wenbin An (Xi'an Jiaotong University), Shijian Lu (Nanyang Technological University)
CodeObject DetectionTransformerVision Language ModelImageMultimodality
π― What it does: This study investigates the object hallucination problem in large visual language models (LVLM) and proposes a training-free, pluggable decoding methodβAssembly of Global and Local Attention (AGLA). This method reduces hallucinations and enhances visual understanding by combining the global features of the original image with the local features of the augmented image.
MLVU: Benchmarking Multi-task Long Video Understanding
Junjie Zhou (Beijing University of Posts and Telecommunications), Zheng Liu (Beijing University of Posts and Telecommunications)
CodeGenerationTransformerLarge Language ModelVideoMultimodalityBenchmark
π― What it does: This paper proposes a new long video understanding benchmark, MLVU, for systematically evaluating the multi-task performance of multimodal large language models (MLLMs) on long videos, and experiments were conducted on 23 of the latest MLLMs.
CodeDomain AdaptationRepresentation LearningTransformerVision Language ModelContrastive LearningImageMultimodality
π― What it does: This paper proposes the MMRL framework, which introduces a shared learnable multimodal representation space on the pre-trained CLIP VLM, and inserts representation subwords in the high-level encoder, achieving efficient adaptation to downstream tasks with a small number of samples while maintaining generality.
CodeOptimizationRobotic IntelligenceSimultaneous Localization and MappingOptical FlowPoint CloudMesh
π― What it does: This paper proposes MNE-SLAM, a distributed neural SLAM system that supports multi-robot collaboration, capable of performing distributed mapping, camera tracking, loop closure, and subgraph fusion without sharing raw sensor data.
π― What it does: A lightweight visual network called MobileMamba is proposed and implemented, integrating multi-scale perception and high-frequency detail extraction, suitable for downstream tasks such as classification, detection, and segmentation at high resolutions.
π― What it does: Proposes the AMNAR framework, which utilizes task graphs and dynamic programming to predict all valid next actions, and dynamically reconstructs normative action representations for each action to detect erroneous behaviors in process videos.
Modeling Thousands of Human Annotators for Generalizable Text-to-Image Person Re-identification
Jiayu Jiang (South China University of Technology), Xiangmin Xu (South China University of Technology)
CodeRetrievalTransformerLarge Language ModelPrompt EngineeringImageText
π― What it does: Developed the Human Annotator Modeling (HAM) method, which utilizes multimodal large language models (MLLM) to automatically generate diverse text descriptions, thereby enhancing human retrieval performance from text to image.
Mono-InternVL: Pushing the Boundaries of Monolithic Multimodal Large Language Models with Endogenous Visual Pre-training
Gen Luo (OpenGVLab Shanghai AI Laboratory), Xizhou Zhu (Tsinghua University)
CodeTransformerLarge Language ModelMixture of ExpertsVision Language ModelTextMultimodality
π― What it does: Designed and trained Mono-InternVLβa single-modal multimodal model that embeds visual experts into a pre-trained large language model (LLM) and implements a mixture of experts (MoE) structure, proposing an Endogenous Visual Pre-training (EViP) phased coarse-to-fine visual learning process;
Mono3DVLT: Monocular-Video-Based 3D Visual Language Tracking
Hongkai Wei (Chang'an University), Ajmal Saeed Mian (University of Western Australia)
CodeObject DetectionObject TrackingTransformerVision Language ModelVideoTextMultimodality
π― What it does: This paper proposes a monocular video 3D object tracking task guided by natural language, called Mono3DVLT, and provides a complete end-to-end solution.
π― What it does: A monocular 3D object detection framework called MonoDGP based on Transformer is proposed, which utilizes geometric error priors, decoupled queries, and a region segmentation head to achieve more accurate depth estimation and 3D object localization.
π― What it does: This paper proposes MonoSplat, a method for generalizable 3D Gaussian dispersion reconstruction that utilizes knowledge from a pre-trained monocular depth model.
π― What it does: A teaching assistant knowledge distillation framework for monocular 3D detection, MonoTAKD, is proposed, which achieves efficient knowledge transfer under the same modality using a camera-based teaching assistant model and extracts unique 3D spatial information from LiDAR through residual distillation.
π― What it does: This paper proposes a dual-branch structure called MonSter, which combines monocular depth estimation with stereo matching to form an iterative complementary enhancement process.
MOS-Attack: A Scalable Multi-objective Adversarial Attack Framework
Ping Guo (City University of Hong Kong), Zhenkun Wang (Southern University of Science and Technology)
CodeOptimizationAdversarial AttackImage
π― What it does: This paper proposes a multi-objective attack framework called MOS-Attack, which generates stronger adversarial samples by simultaneously optimizing various surrogate loss functions.
π― What it does: This paper proposes a framework called MOS for General Category Discovery (GCD) that utilizes scene information to address the misjudgment problem caused by the confusion between scenes and objects.