CVPR 2024 Papers — Page 14

LayoutFormer: Hierarchical Text Detection Towards Scene Text Understanding

Min Liang (University of Science and Technology Beijing), Xu-Cheng Yin (University of Science and Technology Beijing)

🎯 What it does: A hierarchical text detection framework named LayoutFormer is proposed, which can simultaneously detect and output word, text line, and paragraph-level text boxes along with their geometric layouts, directly supporting hierarchical understanding of scene text.

LayoutLLM: Layout Instruction Tuning with Large Language Models for Document Understanding

Chuwei Luo (Alibaba Group), Cong Yao (Alibaba Group)

🎯 What it does: This paper presents LayoutLLM, a framework for LLM/MLLM fine-tuning through layout instructions to enhance zero-shot performance in document understanding.

LDP: Language-driven Dual-Pixel Image Defocus Deblurring Network

Hao Yang (Beijing Institute of Technology), Miaomiao Liu (Australian National University)

🎯 What it does: Utilize CLIP for unsupervised estimation of the blur map of dual-pixel (DP) images and achieve deblurring through a blur prior attention network.

LEAD: Exploring Logit Space Evolution for Model Selection

Zixuan Hu (Peking University), Ling-Yu Duan (Peking University)

🎯 What it does: This study addresses the model selection problem and proposes the LEAD metric based on logit space evolution, using a first-order differential equation to predict the final performance after fine-tuning in a closed form.

LEAD: Learning Decomposition for Source-free Universal Domain Adaptation

Sanqing Qu (Tongji University), Changjun Jiang (Tongji University)

🎯 What it does: The LEAD framework is proposed, which uses source model feature decomposition to identify private categories in the target domain, achieving universal domain adaptation without source data.

Leak and Learn: An Attacker's Cookbook to Train Using Leaked Data from Federated Learning

Joshua C. Zhao (Purdue University), Saurabh Bagchi (Purdue University)

🎯 What it does: This study investigates the impact of data reconstruction attacks on subsequent model training in federated learning, and evaluates the effects of two types of attacks: gradient inversion and linear layer leakage during model training.

LEAP-VO: Long-term Effective Any Point Tracking for Visual Odometry

Weirong Chen (TU Munich), Marc Pollefeys (Microsoft)

🎯 What it does: A Long-Period Effective Arbitrary Point Tracking (LEAP) module and a vision odometry based on it (LEAP-VO) are proposed for robust visual localization in dynamic scenes.

Learn from View Correlation: An Anchor Enhancement Strategy for Multi-view Clustering

Suyuan Liu (National University of Defense Technology), Xinwang Liu (National University of Defense Technology)

🎯 What it does: An anchor point enhancement strategy based on view correlation, AEVC, is proposed, which can improve the quality of anchor points in multi-view clustering and enhance anchor graph construction.

Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation

Jingyun Wang (Beihang University), Guoliang Kang (Beihang University)

🎯 What it does: Utilizing CLIP for unsupervised semantic segmentation, explicitly modeling and correcting its spatial and category biases.

Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans

Romain Loiseau (University Gustave Eiffel), Loic Landrieu (University Gustave Eiffel)

🎯 What it does: An unsupervised method is proposed that utilizes learnable 3D prototypes and a probabilistic slot model to decompose large-scale aerial LiDAR point clouds into interpretable geometric parts for instance and semantic segmentation.

Learned Lossless Image Compression based on Bit Plane Slicing

Zhe Zhang (Wuhan University), Shan Liu (Tencent Media Lab)

🎯 What it does: The ArIB-BPS framework is proposed, achieving lossless image compression by slicing the bit plane and combining hierarchical latent variables with sub-image autoregression.

Learned Representation-Guided Diffusion Models for Large-Image Generation

Alexandros Graikos (Stony Brook University), Dimitris Samaras (Stony Brook University)

🎯 What it does: This paper proposes the use of self-supervised learning (SSL) features as conditions for diffusion models to achieve high-quality block-level generation of pathological and remote sensing images, and to generate large images by stitching conditional image blocks together; it also uses generated images for data augmentation and explores text-to-large-image generation schemes.

Learned Scanpaths Aid Blind Panoramic Video Quality Assessment

Kanglong Fan (City University of Hong Kong), Kede Ma (City University of Hong Kong)

🎯 What it does: An end-to-end blind panoramic video quality assessment method is proposed, utilizing a differentiable scanning path generator and quality assessor trained jointly.

Learned Trajectory Embedding for Subspace Clustering

Yaroslava Lochman (Chalmers University of Technology), Christopher Zach (Lund University)

🎯 What it does: A method is proposed to map point trajectories to embedding vectors for subspace clustering and trajectory completion;

Learning Adaptive Spatial Coherent Correlations for Speech-Preserving Facial Expression Manipulation

Tianshui Chen (Guangdong University of Technology), Liang Lin (Sun Yat-Sen University)

🎯 What it does: This paper proposes an Adaptive Spatial Consistency Correlation Learning (ASCCL) framework for Speech-Driven Facial Expression Manipulation (SPFEM), which learns the high correlation of local facial animations of the same spoken content under different emotions as additional supervision to guide the expression generation model.

Learning Background Prompts to Discover Implicit Knowledge for Open Vocabulary Object Detection

Jiaming Li (Sun Yat-sen University), Guanbin Li (Sun Yat-sen University)

🎯 What it does: This paper proposes an open vocabulary object detection framework named LBP, which enhances the detection performance of base and novel classes by learning background prompts to mine and utilize the implicit object knowledge in the background.

Learning by Correction: Efficient Tuning Task for Zero-Shot Generative Vision-Language Reasoning

Rongjie Li (ShanghaiTech University), Xuming He (ShanghaiTech University)

🎯 What it does: This paper proposes an image conditional caption correction task (ICCC) that does not require manual annotation. It enhances the generative performance of visual language models (VLM) in zero-shot visual language reasoning tasks by automatically correcting errors when there is a mismatch between images and text.

Learning CNN on ViT: A Hybrid Model to Explicitly Class-specific Boundaries for Domain Adaptation

Ba Hung Ngo (Chonnam National University), Tae Jong Choi

🎯 What it does: A hybrid model ECB that integrates ViT and CNN is proposed, utilizing ViT to extract global features and CNN to cluster local features, achieving domain adaptation.

Learning Continual Compatible Representation for Re-indexing Free Lifelong Person Re-identification

Zhenyu Cui (Peking University), Yuxin Peng (Peking University)

🎯 What it does: This paper proposes a task of lifelong person ReID (RFL-ReID) that achieves this without re-indexing the original images, and introduces the Continual Compatible Representation (C2R) method.

Learning Continuous 3D Words for Text-to-Image Generation

Ta-Ying Cheng (University of Oxford), Niki Trigoni (University of Oxford)

🎯 What it does: By learning Continuous 3D Words, the text-to-image model achieves fine-grained control over continuous attributes such as lighting, pose, and camera parameters based on a single or a few 3D meshes.

Learning Correlation Structures for Vision Transformers

Manjin Kim (POSTECH), Minsu Cho (POSTECH)

🎯 What it does: A structural self-attention mechanism, StructSA, is proposed, and based on this, the StructViT model is constructed to enhance the performance of visual Transformers in image and video classification tasks.

Learning Coupled Dictionaries from Unpaired Data for Image Super-Resolution

Longguang Wang (Aviation University of Air Force), Yulan Guo

🎯 What it does: Proposes an image super-resolution method based on learning coupled dictionaries from unpaired data.

Learning Degradation-Independent Representations for Camera ISP Pipelines

Yanhui Guo (McMaster University), Xiaolin Wu (McMaster University)

🎯 What it does: A self-supervised learning method is proposed to generate image representations (DiR) that are independent of camera ISP degradation, and this representation is used to enhance image restoration and downstream task performance.

Learning Degradation-unaware Representation with Prior-based Latent Transformations for Blind Face Restoration

Lianxin Xie (South China University of Technology), Hau San Wong (City University of Hong Kong)

🎯 What it does: A prior-based Potential Transformation (PLTrans) framework is proposed for blind face restoration.

Learning Diffusion Texture Priors for Image Restoration

Tian Ye (Hong Kong University of Science and Technology), Lei Zhu (Hong Kong University of Science and Technology)

🎯 What it does: This paper proposes a texture prior framework based on diffusion models, DTPM, which freezes most parameters using a pre-trained large-scale high-quality texture prior in downstream image restoration tasks. It quickly fine-tunes through an initial predictor and a conditional adapter, capable of generating high-quality details in various restoration tasks (deblurring, dehazing/de-raining, snow removal, raindrop removal).

Learning Discriminative Dynamics with Label Corruption for Noisy Label Detection

Suyeon Kim (POSTECH), Hwanjo Yu (POSTECH)

🎯 What it does: The DynaCor framework is proposed, which combines training dynamics and label corruption to train models on both original and artificially noisy data, generating dynamic trajectories. A dynamic encoder is then used to cluster and distinguish clean samples from noisy samples.

Learning Disentangled Identifiers for Action-Customized Text-to-Image Generation

Siteng Huang (Zhejiang University), Donglin Wang (Westlake University)

🎯 What it does: For the task of action customization in text-to-image generation, a method for learning action-specific identifiers based on Layer-wise Identifier and Gradient Masking, called ADI, is proposed.

Learning Dynamic Tetrahedra for High-Quality Talking Head Synthesis

Zicheng Zhang (University of Chinese Academy of Sciences), Ming Yang (Ant Group)

🎯 What it does: A hybrid representation called DynTet is proposed, which combines neural networks with tetrahedral meshes for high-quality, real-time speaker head synthesis.

Learning Equi-angular Representations for Online Continual Learning

Minhyuk Seo (Yonsei University), Jonghyun Choi (Seoul National University)

🎯 What it does: An online continual learning method called EARL is proposed, achieving efficient learning through isometric representation learning under single-round training.

Learning for Transductive Threshold Calibration in Open-World Recognition

Qin Zhang (Amazon Web Services), Yifan Xing (Amazon Web Services)

🎯 What it does: A threshold calibration task is proposed in open-world visual recognition, addressing the need for thresholding to achieve the desired TPR or TNR for embedding models trained on closed sets during open-class testing.

Learning from Observer Gaze: Zero-Shot Attention Prediction Oriented by Human-Object Interaction Recognition

Yuchen Zhou (Sun Yat-sen University), Chao Gou (Sun Yat-sen University)

🎯 What it does: This paper studies interaction-oriented visual attention prediction, proposing the Zero-shot Interaction Attention task (ZeroIA) and designing the Interaction Attention model (IA), while also constructing the first large-scale interactive eye-tracking dataset IG.

Learning from One Continuous Video Stream

João Carreira (Google DeepMind), Andrew Zisserman (University of Oxford)

🎯 What it does: This study investigates an online learning framework under a single continuous video stream and evaluates the model's adaptability and generalization performance.

Learning from Synthetic Human Group Activities

Che-Jui Chang (Rutgers University), Mubbasir Kapadia (Roblox)

🎯 What it does: This paper presents M Act 3—a synthetic data generator based on the Unity Engine, capable of generating multi-view, multi-group, and multi-person human atomic actions and group activity data, along with rich 2D/3D annotations;

Learning Group Activity Features Through Person Attribute Prediction

Chihiro Nakatani (Toyota Technological Institute), Norimichi Ukita (Toyota Technological Institute)

🎯 What it does: Learn multi-person crowd activity features (Group Activity Feature, GAF) using only human attributes (action categories or appearance features) instead of group activity labels, and utilize this feature for attribute prediction and retrieval.

Learning Inclusion Matching for Animation Paint Bucket Colorization

Yuekun Dai (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)

🎯 What it does: Proposes an animation brush coloring automation process based on inclusive matching, achieving for the first time the ability to automatically propagate color to subsequent frames by coloring just one frame.

Learning Instance-Aware Correspondences for Robust Multi-Instance Point Cloud Registration in Cluttered Scenes

Zhiyuan Yu (National University of Defense Technology), Kai Xu (National University of Defense Technology)

🎯 What it does: For multi-instance point cloud registration, this paper proposes a coarse-to-fine instance-aware matcher MIRETR, which directly extracts instance-level correspondences from the scene point cloud and estimates transformations.

Learning Intra-view and Cross-view Geometric Knowledge for Stereo Matching

Rui Gong (Nanyang Technological University), Jun Cheng (Institute for Infocomm Research A*STAR)

🎯 What it does: This paper proposes an ICGNet framework that introduces intra-view and inter-view geometric knowledge through interest point detectors and matchers to enhance stereo matching accuracy.

Learning Large-Factor EM Image Super-Resolution with Generative Priors

Jiateng Shou (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)

🎯 What it does: A generative learning framework for super-resolution of high-magnification electron microscope images (8×, 16×) is proposed, capable of restoring high-quality images at low sampling rates and improving downstream segmentation accuracy.

Learning Multi-Dimensional Human Preference for Text-to-Image Generation

Sixian Zhang (Kuaishou Technology), Zhongyuan Wang (Kuaishou Technology)

🎯 What it does: This paper proposes a multidimensional human preference dataset (MHP) and trains a multidimensional preference scoring model (MPS) based on this dataset to evaluate the outputs of text-to-image generation models across four dimensions (aesthetics, detail, semantic consistency, and overall).

Learning Object State Changes in Videos: An Open-World Perspective

Zihui Xue (University of Texas at Austin), Kristen Grauman (Meta)

🎯 What it does: This paper studies the open-world localization problem of object state changes (OSC) in videos and proposes an end-to-end VIDOSC framework.

Learning Occupancy for Monocular 3D Object Detection

Liang Peng (Zhejiang University), Deng Cai (Zhejiang University)

🎯 What it does: This paper proposes to learn occupancy in both the frustum space and 3D space simultaneously, enhancing the 3D object detection features under a monocular camera, ultimately achieving more accurate 3D detection.

Learning SO(3)-Invariant Semantic Correspondence via Local Shape Transform

Chunghyun Park (POSTECH), Minsu Cho (POSTECH)

🎯 What it does: A self-supervised 3D shape correspondence learning method called RIST is proposed, which achieves semantic correspondence under arbitrary rotations through SO(3)-invariant local shape transformations.

Learning Spatial Adaptation and Temporal Coherence in Diffusion Models for Video Super-Resolution

Zhikai Chen (University of Science and Technology of China), Tao Mei (University of Rochester)

🎯 What it does: A video super-resolution method based on diffusion models, SATeCo, is proposed, which enhances video super-resolution quality using spatial adaptation and temporal consistency modules.

Learning Spatial Features from Audio-Visual Correspondence in Egocentric Videos

Sagnik Majumder (University of Texas at Austin), Kristen Grauman (University of Texas at Austin)

🎯 What it does: Utilizing a self-supervised masked autoencoder framework to learn spatial correspondences from first-person videos and binaural audio, and obtaining audio-visual spatial features through audio-filled pre-training.

Learning Structure-from-Motion with Graph Attention Networks

Lucas Brynte (Chalmers University of Technology), Fredrik Kahl (Chalmers University of Technology)

🎯 What it does: A method for learning structure from motion (SfM) without initialization based on graph attention networks is proposed, which can directly predict camera poses and 3D point coordinates from multi-view 2D keypoints.

Learning the 3D Fauna of the Web

Zizhang Li (Stanford University), Jiajun Wu (Stanford University)

🎯 What it does: Learn to unsupervisedly reconstruct deformable 3D models of over 100 types of quadrupedal animals from internet images.

Learning to Control Camera Exposure via Reinforcement Learning

Kyunghyun Lee (LG AI Research), Byeong-Uk Lee (KRAFTON)

🎯 What it does: A joint exposure parameter control framework based on deep reinforcement learning (DRL-AE) is proposed, achieving real-time and rapid adjustment of camera exposure time and gain.

Learning to Count without Annotations

Lukas Knobel (University of Amsterdam), Yuki M. Asano (University of Amsterdam)

🎯 What it does: The UnCounTR model is proposed, which learns reference-based counting tasks under completely unsupervised conditions using self-supervised generated Self-Collages.

Learning to Localize Objects Improves Spatial Reasoning in Visual-LLMs

Kanchana Ranasinghe (Stony Brook University), Tsung-Yu Lin (Meta)

🎯 What it does: This paper proposes LocVLM by incorporating coordinate description into instruction fine-tuning in visual-LLM to enhance the model's spatial reasoning and localization capabilities.

Learning to Navigate Efficiently and Precisely in Real Environments

Guillaume Bono (NAVER LABS Europe), Christian Wolf (NAVER LABS Europe)

🎯 What it does: An end-to-end visual navigation agent capable of efficient and accurate navigation was designed and trained in a real indoor environment. During training, a second-order dynamic model identified from real robot data was incorporated into the Habitat simulator, along with noisy odometry and localization information.

Learning to Predict Activity Progress by Self-Supervised Video Alignment

Gerard Donahue (Northeastern University), Ehsan Elhamifar (Northeastern University)

🎯 What it does: This paper proposes a self-supervised video alignment and progress prediction framework called GTCC, which addresses issues of uncertain action sequences, repetitive actions, and background frames.

Learning to Produce Semi-dense Correspondences for Visual Localization

Khang Truong Giang (KAIST), Sungho Jo (KAIST)

🎯 What it does: A visual localization method based on semi-dense 2D-2D matching is proposed, which maps all detected 2D keypoints to 3D space through a Point Inference Network, generating a large number of 2D-3D correspondences.

Learning to Rank Patches for Unbiased Image Redundancy Reduction

Yang Luo (Fudan University), Yu-Gang Jiang (Fudan University)

🎯 What it does: A self-supervised image redundancy reduction framework LTRP has been developed, utilizing MAE to reconstruct differences and generate patch importance pseudo-labels, and selecting high-information image blocks through learned ranking.

Learning to Rematch Mismatched Pairs for Robust Cross-Modal Retrieval

Haochen Han (Xi'an Jiaotong University), Jingdong Wang (Baidu Inc)

🎯 What it does: This paper addresses the issue of Partial Mismatched Pairs (PMPs) in cross-modal retrieval and proposes a framework called L2RM based on Optimal Transport, which enhances retrieval robustness by learning to rematch mismatched samples.

Learning to Remove Wrinkled Transparent Film with Polarized Prior

Jiaqi Tang (Hong Kong University of Science and Technology), Ying-Cong Chen (Hong Kong University of Science and Technology)

🎯 What it does: This paper studies a new industrial image processing task - removing wrinkled transparent films to recover the image information obscured by the film.

Learning to Segment Referred Objects from Narrated Egocentric Videos

Yuhan Shen (Meta), Effrosyni Mavroudi (Meta)

🎯 What it does: This paper proposes a weakly supervised narrative-driven video object segmentation task (NVOS) and implements the ROSA framework, which can achieve pixel-level segmentation of objects mentioned in videos using only narrative text aligned with the video, without the need for pixel-level annotations.

Learning to Select Views for Efficient Multi-View Understanding

Yunzhong Hou (Australian National University), Liang Zheng (Australian National University)

🎯 What it does: This paper proposes a framework for selective view pipelining, which achieves efficient inference by dynamically selecting only 2-3 views in multi-view classification and detection.

Learning to Transform Dynamically for Better Adversarial Transferability

Rongyi Zhu (University of Rochester), Chenliang Xu (University of Rochester)

🎯 What it does: Enhancing the transferability of gradient-based adversarial attacks across multiple models through dynamic learning of input transformation strategies.

Learning to Visually Localize Sound Sources from Mixtures without Prior Source Knowledge

Dongjin Kim (Kyung Hee University), Jung Uk Kim (Kyung Hee University)

🎯 What it does: This study investigates a multi-source audio localization method that utilizes visual-audio collaboration without requiring prior source number information.

Learning Transferable Negative Prompts for Out-of-Distribution Detection

Tianqi Li (Beihang University), Jin Zheng (Beihang University)

🎯 What it does: A method called NegPrompt is proposed, which achieves unsupervised out-of-distribution (OOD) detection by learning negative prompts, utilizing the text-image alignment space of CLIP and relying solely on ID sample training.

Learning Triangular Distribution in Visual World

Ping Chen (MicroBT Inc), Yanlin Qian (Waseda University)

🎯 What it does: The Triangular Distribution Transform (TDT) is proposed as a non-parametric plug-in that converts the nonlinear features extracted by CNN into features that linearly correspond to label differences, allowing regression tasks to be completed using only a linear head.

Learning Vision from Models Rivals Learning Vision from Data

Yonglong Tian (Google Research), Phillip Isola (Massachusetts Institute of Technology)

🎯 What it does: This paper proposes SynCLR, a method for visual representation learning that uses only synthetic text and images, without relying on any real data.

Learning Visual Prompt for Gait Recognition

Kang Ma (Beijing Institute of Technology), Dezhi Zheng (Beijing Institute of Technology)

🎯 What it does: A Visual Prompt Network (VPNet) is proposed to extract global motion patterns and achieve gait recognition in gait sequences through a trainable prompt pool and dynamic Transformer.

Learning with Structural Labels for Learning with Noisy Labels

Noo-ri Kim (Sungkyunkwan University), Jee-Hyong Lee (Sungkyunkwan University)

🎯 What it does: This paper proposes a learning framework (LSL) that utilizes inverse k-nearest neighbors to estimate structural labels, aimed at improving the performance of deep learning models in noisy label environments.

Learning with Unreliability: Fast Few-shot Voxel Radiance Fields with Relative Geometric Consistency

Yingjie Xu (Singapore Management University), Shengfeng He (Singapore Management University)

🎯 What it does: This paper proposes a method for fast sparse NeRF reconstruction using unreliable regions of pseudo-views from a limited number of perspectives.

Learning without Exact Guidance: Updating Large-scale High-resolution Land Cover Maps from Low-resolution Historical Labels

Zhuohong Li (Wuhan University), Hongyan Zhang (China University of Geosciences)

🎯 What it does: An end-to-end weakly supervised framework called Paraformer is proposed, which utilizes low-resolution historical labels to update large-scale high-resolution land cover maps.

LED: A Large-scale Real-world Paired Dataset for Event Camera Denoising

Yuxing Duan (Huazhong University of Science and Technology)

🎯 What it does: A large-scale real event denoising dataset LED has been constructed, and a dual event denoising framework DED and a dynamic threshold LIF neuron-based SNN denoising model DTSNN have been proposed.

LEDITS++: Limitless Image Editing using Text-to-Image Models

Manuel Brack, Apolinario Passos

🎯 What it does: This paper presents LEDITS++, a method for multi-faceted, precise, and controllable text editing of real images that does not require fine-tuning.

LeftRefill: Filling Right Canvas based on Left Reference through Generalized Text-to-Image Diffusion Model

Chenjie Cao (Fudan University), Yanwei Fu (Fudan University)

🎯 What it does: This paper proposes LeftRefill, a general framework for end-to-end reference-guided image inpainting (Ref-inpainting) and novel view synthesis (NVS) by horizontally stitching reference and target images, utilizing prompt tuning and self-attention.

LeGO: Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example

Soyeon Yoon (Korea Advanced Institute of Science and Technology), Junyong Noh (Korea Advanced Institute of Science and Technology)

🎯 What it does: Utilize a surface deformation network and a single example to generate animatable stylized 3D face models;

LEMON: Learning 3D Human-Object Interaction Relation from 2D Images

Yuhang Yang, Zheng-Jun Zha

🎯 What it does: Proposes the LEMON method, which jointly predicts human contact, object affordance, and spatial relationships in 3D human-computer interaction.

LEOD: Label-Efficient Object Detection for Event Cameras

Ziyi Wu (University of Toronto), Igor Gilitschenski (University of Zurich)

🎯 What it does: Proposes the LEOD framework to achieve efficient target detection with event cameras.

Let's Think Outside the Box: Exploring Leap-of-Thought in Large Language Models with Creative Humor Generation

Shanshan Zhong (Sun Yat-Sen University), Pan Zhou (Singapore Management University)

🎯 What it does: The Creative Leap-of-Thought (CLoT) framework is proposed to enhance the Leap-of-Thought (non-sequential, associative creative reasoning ability) of large language models, primarily used for multimodal humor generation (Oogiri game).

Leveraging Camera Triplets for Efficient and Accurate Structure-from-Motion

Lalit Manam (Indian Institute of Science), Venu Madhav Govindu (Indian Institute of Science)

🎯 What it does: This paper proposes an edge scoring mechanism based on camera triplets to simultaneously achieve the sparsification of the view graph and the debiasing of repeated structures in Structure from Motion (SfM).

Leveraging Cross-Modal Neighbor Representation for Improved CLIP Classification

Chao Yi (Nanjing University), Han-Jia Ye (Nanjing University)

🎯 What it does: This paper proposes a cross-modal neighbor representation based on CLIP image-text distance (CODER) and constructs a high-quality neighbor text set by generating diverse texts (ATG) through a large language model, thereby improving CLIP's feature extraction and prediction in zero-shot and few-shot image classification tasks.

Leveraging Frame Affinity for sRGB-to-RAW Video De-rendering

Chen Zhang (SenseTime Research and Tetras.AI), Wentao Liu (SenseTime Research and Tetras.AI)

🎯 What it does: A frame affinity-based sRGB-to-RAW video de-rendering framework is proposed, which only requires saving one frame of RAW as a prior during capture to recover the entire RAW video.

Leveraging Predicate and Triplet Learning for Scene Graph Generation

Jiankai Li (Beihang University), Weixin Li (Beihang University)

🎯 What it does: A Dual-Granularity Relation Modeling Network (DRM) is proposed, which simultaneously utilizes coarse-grained predicate features and fine-grained triplet features, and alleviates the long-tail problem through Dual-Granularity Knowledge Transfer (DKT) to improve scene graph generation accuracy.

Leveraging Vision-Language Models for Improving Domain Generalization in Image Classification

Sravanti Addepalli (Indian Institute of Science), R. Venkatesh Babu (Indian Institute of Science)

🎯 What it does: A black-box knowledge distillation method (VL2V-ADiP) is proposed, which aligns and distills the visual and language representations in Vision-Language models (such as CLIP) to a pre-trained single-modal visual student model, enhancing the domain generalization ability of image classification; simultaneously, a self-distillation method (VL2V-SD) is proposed in a white-box scenario to enhance the visual representation of VLM.

LiDAR-based Person Re-identification

Wenxuan Guo (Tsinghua University), Jie Zhou (Tsinghua University)

🎯 What it does: The study utilizes LiDAR point clouds for person re-identification, proposing the ReID3D framework and constructing two datasets: LReID and LReID-sync.

LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor Scenes

Yanwen Guo (Nanjing University), Hongyu Liu (Nanjing University)

🎯 What it does: A large-scale indoor real LiDAR scanning point cloud dataset LiDAR-Net has been constructed, containing 360 million points, covering 30,000 m² of learning, working, and living scenes, and providing semantic, instance, and detection labels for each raw point; meanwhile, fine-grained annotations for scanning anomalies (reflection noise, ghost points) are provided.

LiDAR4D: Dynamic Neural Fields for Novel Space-time View LiDAR Synthesis

Zehan Zheng (Tongji University), Changjun Jiang (Tongji University)

🎯 What it does: This paper proposes the LiDAR4D framework, which utilizes neural fields to achieve space-time view synthesis of LiDAR point clouds in dynamic scenes.

LidaRF: Delving into Lidar for Neural Radiance Field on Street Scenes

Shanlin Sun (University of California), Manmohan Chandraker (NEC Labs America)

🎯 What it does: This paper proposes a method to enhance the rendering quality of street scene NeRF using LiDAR data, primarily achieved through the integration of LiDAR features, robust occlusion-aware depth supervision, and LiDAR-based viewpoint augmentation.

Lift3D: Zero-Shot Lifting of Any 2D Vision Model to 3D

Mukund Varma T (University of California San Diego), Ravi Ramamoorthi (University of California San Diego)

🎯 What it does: This paper presents Lift3D, which can elevate any pre-trained 2D visual model to 3D without the need for additional training, generating perspective-consistent predictions.

Light the Night: A Multi-Condition Diffusion Framework for Unpaired Low-Light Enhancement in Autonomous Driving

Jinlong Li (Cleveland State University), Hongkai Yu (Cleveland State University)

🎯 What it does: A multi-conditional framework called LightDiff based on diffusion models has been developed for augmenting low-light images in autonomous driving scenarios without paired data, enhancing perception performance.

LightIt: Illumination Modeling and Control for Diffusion Models

Peter Kocsis (Technical University of Munich), Yannick Hold-Geoffroy (Adobe Research)

🎯 What it does: This paper proposes LightIt, an explicit lighting control method based on diffusion models, which utilizes direct shadows and normal information for text-driven image generation and relighting.

LightOctree: Lightweight 3D Spatially-Coherent Indoor Lighting Estimation

Xuecan Wang (Beihang University), Xiaohui Liang (Zhongguancun Laboratory)

🎯 What it does: Achieve 3D spatially consistent estimation of indoor lighting using a single RGB image through a sparse voxel octree network.

Linguistic-Aware Patch Slimming Framework for Fine-grained Cross-Modal Alignment

Zheren Fu (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)

🎯 What it does: A language-aware Patch pruning framework LAPS is designed to improve the cross-modal fine-grained alignment of pure Transformers.

Link-Context Learning for Multimodal LLMs

Yan Tai (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)

🎯 What it does: A framework called 'Link-Context Learning (LCL)' is proposed to enhance the ability of multimodal large language models to learn through causal associations and retain new concepts in dialogues; based on this, Shikra is fine-tuned, and the ISEKAI dataset is constructed for evaluation.

LION: Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge

Gongwei Chen (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)

🎯 What it does: LION enhances multimodal large language models through a dual-layer visual knowledge approach, thereby improving image understanding and reasoning capabilities.

LiSA: LiDAR Localization with Semantic Awareness

Bochun Yang (Xiamen University), Cheng Wang (Xiamen University)

🎯 What it does: This study proposes LiSA, which utilizes semantic information to enhance the Scene Coordinate Regression (SCR) method for LiDAR point cloud localization.

LISA: Reasoning Segmentation via Large Language Model

Xin Lai (Chinese University of Hong Kong), Jiaya Jia (Chinese University of Hong Kong)

LiveHPS: LiDAR-based Scene-level Human Pose and Shape Estimation in Free Environment

Yiming Ren (ShanghaiTech University), Yuexin Ma (ShanghaiTech University)

🎯 What it does: This paper proposes the LiveHPS method under a single LiDAR sensor for capturing human 3D posture, shape, and global translation parameters in large-scale seamless environments.

Living Scenes: Multi-object Relocalization and Reconstruction in Changing 3D Environments

Liyuan Zhu (Stanford University), Iro Armeni (Stanford University)

🎯 What it does: This paper proposes a unified framework named MORE 2, which can perform instance matching, registration, and reconstruction in sparse multi-temporal scans, constructing a long-term cumulative model of dynamic 3D environments.

LL3DA: Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning

Sijin Chen (Fudan University), Tao Chen (Fudan University)

🎯 What it does: A 3D large language assistant LL3DA is proposed, which can simultaneously accept text instructions and visual interactions (such as clicks and box annotations) to understand, reason, and plan in complex 3D scenes.

LLaFS: When Large Language Models Meet Few-Shot Segmentation

Lanyun Zhu (Singapore University of Technology and Design), Jun Liu (Singapore University of Technology and Design)

🎯 What it does: Proposes the LLaFS framework, which utilizes large language models for few-shot image segmentation, achieving an end-to-end inference process.

LLaMA-Excitor: General Instruction Tuning via Indirect Feature Interaction

Bo Zou (Tsinghua University), Youjian Zhao (Tsinghua University)

🎯 What it does: This paper presents LLaMA-Excitor, a PEFT method based on indirect feature interaction. By inserting learnable Excitor blocks into the attention layers of LLaMA, it achieves efficient fine-tuning for instruction following and multimodal tasks while maintaining the original model capabilities.

LLM4SGG: Large Language Models for Weakly Supervised Scene Graph Generation

Kibum Kim (Korea Advanced Institute of Science and Technology), Chanyoung Park (Korea Advanced Institute of Science and Technology)

🎯 What it does: This paper proposes a method for weakly supervised scene graph generation (LLM4SGG) using large language models (LLM), leveraging Chain-of-Thought and in-context few-shot prompting in the steps of triplet extraction and entity/predicate alignment, addressing the issues of semantic oversimplification caused by traditional rule parsers and low density due to dictionary matching.

LLMs are Good Action Recognizers

Haoxuan Qu (Singapore University of Technology and Design), Jun Liu (Singapore University of Technology and Design)

🎯 What it does: Using large language models as action recognizers, first discretizing skeletal sequences into 'action sentences' with VQ-VAE, and then classifying them using the pre-trained LLaMA.

LLMs are Good Sign Language Translators

Jia Gong (Singapore University of Technology and Design), Jun Liu (Singapore University of Technology and Design)

🎯 What it does: This paper proposes a sign language translation framework called SignLLM, which maps sign language videos to a linguistic discrete symbol sequence and then uses a large language model (LLM) to generate the corresponding text translation.

LMDrive: Closed-Loop End-to-End Driving with Large Language Models

Hao Shao (Chinese University of Hong Kong MMLab), Hongsheng Li (Chinese University of Hong Kong)

🎯 What it does: We propose LMDrive—a closed-loop end-to-end language-driven autonomous driving framework that can simultaneously receive multimodal sensor data and natural language instructions, and output vehicle control commands in real-time.