ICCV 2023 Papers — Page 11
IEEE/CVF International Conference on Computer Vision · 2156 papers
LaPE: Layer-adaptive Position Embedding for Vision Transformers with Independent Layer Normalization
Runyi Yu (Zhejiang University), Jie Chen (Peking University)
ClassificationObject DetectionSegmentationTransformerImage
🎯 What it does: This paper studies the position embedding problem in visual Transformers and proposes a Layer Adaptive Position Embedding (LaPE) scheme, which uses two independent layer normalizations to process token embeddings and position embeddings separately at each layer, and progressively integrates position embeddings layer by layer.
Large Selective Kernel Network for Remote Sensing Object Detection
Yuxuan Li (Nankai University), Xiang Li (Nankai University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: A lightweight Large Selective Kernel Network (LSKNet) is proposed, which enhances remote sensing object detection by dynamically adjusting large convolution kernels.
Large-Scale Land Cover Mapping with Fine-Grained Classes via Class-Aware Semi-Supervised Semantic Segmentation
Runmin Dong (Tsinghua University), Haohuan Fu (Tsinghua University)
SegmentationConvolutional Neural NetworkContrastive LearningImageAgriculture Related
🎯 What it does: A unified class-aware semi-supervised semantic segmentation framework is proposed for large-scale fine-grained land cover mapping.
Large-Scale Person Detection and Localization Using Overhead Fisheye Cameras
Lu Yang (Beijing University of Posts and Telecommunications), Wenguan Wang (Zhejiang University)
Object DetectionTransformerImageVideo
🎯 What it does: A large overhead fisheye dataset LOAF has been established, and a human detection and localization system based on fisheye cameras has been proposed.
LaRS: A Diverse Panoptic Maritime Obstacle Detection Dataset and Benchmark
Lojze Žust (University of Ljubljana), Matej Kristan (University of Ljubljana)
Object DetectionSegmentationTransformerImageVideoBenchmark
🎯 What it does: This paper proposes and releases the first diversified marine panoramic obstacle detection benchmark, LaRS, which includes over 4000 frames of key frames with pixel-level panoramic annotations, equipped with 9 frames of context and an online evaluation server.
Late Stopping: Avoiding Confidently Learning from Mislabeled Examples
Suqin Yuan (University of Sydney), Tongliang Liu (University of Sydney)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: A new framework for learning with noisy labels is studied, proposing Late Stopping, which delays stopping and uses the First k rounds Learning (FkL) metric to filter noisy samples, retaining more clear hard examples.
Latent-OFER: Detect, Mask, and Reconstruct with Latent Vectors for Occluded Facial Expression Recognition
Isack Lee (Chonnam National University), Seok Bong Yoo (Chonnam National University)
ClassificationRecognitionConvolutional Neural NetworkTransformerImage
🎯 What it does: A three-stage framework for recognizing occluded facial expressions is proposed: first, use ViT-SVDD to detect occluded patches; then, restore the complete face through a hybrid reconstruction network that integrates ViT and CNN; finally, extract expression-related ViT latent vectors from the reconstruction process and perform expression classification together with CNN features.
LATR: 3D Lane Detection from Monocular Images with Transformer
Yueru Luo (Chinese University of Hong Kong Shenzhen), Zhen Li (Chinese University of Hong Kong Shenzhen)
Object DetectionAutonomous DrivingTransformerImage
🎯 What it does: This paper proposes LATR, an end-to-end 3D lane detection framework based on Transformer, which directly performs 3D lane localization on front-view images, eliminating the need for traditional intermediate 3D representations such as BEV or projections.
LAW-Diffusion: Complex Scene Generation by Diffusion with Layouts
Binbin Yang (Sun Yat-Sen University), Liang Lin (Sun Yat-Sen University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A layout-to-image generation framework called LAW-Diffusion based on diffusion models has been designed and implemented, which can accurately generate realistic and semantically consistent complex scene images based on given multi-object layouts.
LayoutDiffusion: Improving Graphic Layout Generation by Discrete Diffusion Probabilistic Models
Junyi Zhang (Shanghai Jiao Tong University), Dongmei Zhang (Microsoft Research Asia)
GenerationTransformerDiffusion modelImage
🎯 What it does: A graphic layout generation method based on discrete diffusion models, LayoutDiffusion, is designed to automatically generate diverse and high-quality graphic layouts.
LD-ZNet: A Latent Diffusion Approach for Text-Based Image Segmentation
Koutilya PNVR (University of Maryland), David Jacobs (University of Maryland)
SegmentationConvolutional Neural NetworkDiffusion modelImageText
🎯 What it does: This paper proposes a text-driven image segmentation method based on latent diffusion models, ZNet and LD-ZNet, which utilizes the compressed latent space z of LDM and internal visual-language features to achieve more robust semantic segmentation.
LDL: Line Distance Functions for Panoramic Localization
Junho Kim (Seoul National University), Young Min Kim (Seoul National University)
Pose EstimationSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: This paper proposes a panoramic image localization method based on line segments, called LDL (Line Distance Functions for Panoramic Localization). By matching distance functions between line segments extracted from a 3D map and a query panoramic image, it quickly obtains a rough pose, which is then refined using local features and PnP-RANSAC.
LDP-Feat: Image Features with Local Differential Privacy
Francesco Pittaluga (NEC Labs America), Bingbing Zhuang (NEC Labs America)
Safty and PrivacyAdversarial AttackImage
🎯 What it does: This paper proposes two types of reverse attacks against existing adversarial affine subspace embeddings and introduces the first privacy-preserving method, LDP-FEAT, which provides theoretical privacy guarantees for image features based on Local Differential Privacy (LDP).
LEA2: A Lightweight Ensemble Adversarial Attack via Non-overlapping Vulnerable Frequency Regions
Yaguan Qian, Bin Wang
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a new method to address a specific problem in computer vision, with specific details not provided.
LeaF: Learning Frames for 4D Point Cloud Sequence Understanding
Yunze Liu (Tsinghua University), Li Yi (Tsinghua University)
RecognitionSegmentationGraph Neural NetworkTransformerPoint Cloud
🎯 What it does: This paper proposes a 4D point cloud sequence learning framework called LeaF, which utilizes a rotation-equivariant network to learn hierarchical region coordinate frames and separates and fuses geometry and motion through frame-guided 4D convolution/Transformer, significantly improving performance on 4D tasks.
Leaping Into Memories: Space-Time Deep Feature Synthesis
Alexandros Stergiou (Vrije Universiteit Brussel), Nikos Deligiannis (Vrije Universiteit Brussel)
RecognitionGenerationData SynthesisConvolutional Neural NetworkTransformerVideo
🎯 What it does: This paper proposes the LEAPS method, which pre-activates a fixed space-time model using video stimuli, and then optimizes it on noisy videos to synthesize visual videos related to the target action.
Learn TAROT with MENTOR: A Meta-Learned Self-Supervised Approach for Trajectory Prediction
Mozhgan Pourkeshavarz (Huawei), Amir Rasouli (Huawei)
Autonomous DrivingMeta LearningGraph Neural NetworkTransformerGraphTime Series
🎯 What it does: This paper proposes a meta-learning-based trajectory prediction framework called MENTOR, which enhances prediction diversity and legality by learning road topology through a self-supervised task called TAROT.
Learned Compressive Representations for Single-Photon 3D Imaging
Felipe Gutierrez-Barragan (University of Wisconsin-Madison), Andreas Velten (University of Wisconsin-Madison)
Depth EstimationCompressionOptimizationConvolutional Neural NetworkImage
🎯 What it does: A linear compression representation that can be constructed in real-time within a SPAD pixel array is proposed for compressing 3D timestamp histograms, with the compressed results subsequently fed into a 3D CNN for depth reconstruction; simultaneously achieving joint optimization of the end-to-end learning of the compression code and depth estimation model.
Learned Image Reasoning Prior Penetrates Deep Unfolding Network for Panchromatic and Multi-spectral Image Fusion
Man Zhou (Nanyang Technological University), Chongyi Li (Nanyang Technological University)
Image TranslationRestorationSuper ResolutionConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: Using the pre-trained MAE as an image inference prior, a deep unfolding network is embedded to achieve high-resolution multispectral image fusion.
Learning a More Continuous Zero Level Set in Unsigned Distance Fields through Level Set Projection
Junsheng Zhou (Tsinghua University), Zhizhong Han
SegmentationGenerationAutonomous DrivingPoint CloudBenchmark
🎯 What it does: This paper proposes the use of hierarchical projection constraints to learn a more continuous and accurate zero isosurface set, thereby enhancing three-dimensional applications such as surface reconstruction, point cloud normal estimation, and upsampling based on unsigned distance fields (UDF).
Learning a Room with the Occ-SDF Hybrid: Signed Distance Function Mingled with Occupancy Aids Scene Representation
Xiaoyang Lyu (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)
GenerationDepth EstimationNeural Radiance FieldPoint CloudMesh
🎯 What it does: This paper proposes a neural field that combines feature rendering loss and occupancy-SDF mixed representation to recover high-detail 3D geometry of indoor scenes from multi-view images.
Learning Adaptive Neighborhoods for Graph Neural Networks
Avishkar Saha (University of Surrey), Richard Bowden (University of Surrey)
ClassificationOptimizationGraph Neural NetworkAuto EncoderPoint CloudGraph
🎯 What it does: A differentiable graph generation module (DGG) is proposed, which can learn adaptive neighborhood sizes and generate graph topologies in any GCN, replacing preset or existing adjacency matrices.
Learning by Sorting: Self-supervised Learning with Group Ordering Constraints
Nina Shvetsova (Goethe University Frankfurt), Hilde Kuehne (Goethe University Frankfurt)
RetrievalRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A self-supervised contrastive learning loss GroCo based on differentiable sorting networks is proposed, treating positive and negative samples as groups and optimizing their sorting constraints.
Learning Clothing and Pose Invariant 3D Shape Representation for Long-Term Person Re-Identification
Feng Liu (Michigan State University), Xiaoming Liu (Michigan State University)
RecognitionRetrievalConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Proposed the 3DInvarReID method, which utilizes a 3D implicit network for long-term person re-identification to achieve clothing and pose invariant 3D shape representation.
Learning Concise and Descriptive Attributes for Visual Recognition
An Yan (University of California San Diego), Julian McAuley (University of California San Diego)
ClassificationRecognitionTransformerLarge Language ModelVision Language ModelImage
🎯 What it does: This paper proposes a new paradigm for visual recognition using a small number of interpretable attributes, automatically extracting attributes from large language models and filtering a refined and discriminative set of attributes through a learning-search method;
Learning Concordant Attention via Target-aware Alignment for Visible-Infrared Person Re-identification
Jianbing Wu (Peking University), Hao Tang (ETH Zurich)
RecognitionRetrievalKnowledge DistillationConvolutional Neural NetworkImageMultimodality
🎯 What it does: This paper proposes a framework called Concordant Attention Learning (CAL) for human re-identification between visible and infrared images.
Learning Continuous Exposure Value Representations for Single-Image HDR Reconstruction
Su-Kai Chen (National Yang Ming Chiao Tung University), Yen-Yu Lin (National Yang Ming Chiao Tung University)
RestorationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This study proposes a Continuous Exposure Value Representation (CEVR) method that uses implicit neural functions to generate LDR images with arbitrary exposure values, thereby constructing a denser and more continuous LDR stack, which is then fused using the Debevec method to obtain high-quality HDR images.
Learning Correction Filter via Degradation-Adaptive Regression for Blind Single Image Super-Resolution
Hongyang Zhou (University of Science and Technology Beijing), Xu-Cheng Yin (University of Science and Technology Beijing)
RestorationSuper ResolutionGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes an unsupervised learning method for degradation correction filtering, utilizing a spatially variable degradation adaptive regression module (DARM) to transform the complex degradation of low-resolution images into known degradation, allowing existing super-resolution networks to operate in blind scenarios.
Learning Cross-Modal Affinity for Referring Video Object Segmentation Targeting Limited Samples
Guanghui Li (Anhui University of Technology), Feng Zheng (Southern University of Science and Technology)
Object DetectionSegmentationTransformerVideoTextBenchmark
🎯 What it does: Proposes a Cross-Modal Affinity (CMA) module to achieve Few-Shot Reference Video Object Segmentation (FS-RVOS) tasks in a few-shot environment.
Learning Cross-Representation Affinity Consistency for Sparsely Supervised Biomedical Instance Segmentation
Xiaoyu Liu (University of Science and Technology of China), Feng Wu (Institute of Artificial Intelligence Hefei Comprehensive National Science Center)
Object DetectionSegmentationConvolutional Neural NetworkBiomedical Data
🎯 What it does: This paper proposes a biomedical instance segmentation framework under sparse instance-level supervision, utilizing two parallel networks to predict implicit affinity graphs and explicit affinity graphs, and training through cross-representation affinity consistency.
Learning Data-Driven Vector-Quantized Degradation Model for Animation Video Super-Resolution
Zixi Tuo (Xi'an Jiaotong University), Xueming Qian
RestorationSuper ResolutionGenerative Adversarial NetworkVideo
🎯 What it does: This paper proposes a multi-scale vector quantization degradation model that utilizes the intrinsic features of animated videos to generate more realistic low-resolution training data, thereby enhancing the super-resolution effect of animated videos.
Learning Depth Estimation for Transparent and Mirror Surfaces
Alex Costanzino (University of Bologna), Luigi Di Stefano (University of Bologna)
Depth EstimationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a simple training strategy for depth estimation, utilizing segmentation masks of transparent/mirror (ToM) objects. After randomly filling these areas with colors, a pre-trained monocular depth network generates 'virtual' depth labels, which are then used to fine-tune the original depth network, thereby improving the depth estimation performance for ToM surfaces.
Learning Fine-Grained Features for Pixel-Wise Video Correspondences
Rui Li (University of Science and Technology of China), Dong Liu (University of Science and Technology of China)
Object TrackingSegmentationDomain AdaptationComputational EfficiencyConvolutional Neural NetworkGenerative Adversarial NetworkContrastive LearningOptical FlowVideo
🎯 What it does: Learning fine-grained features for pixel-level video correspondence, combining synthetic and real video self-supervised training, and employing soft labels, adversarial domain adaptation, and a coarse-to-fine matching framework to improve correspondence accuracy and efficiency.
Learning Foresightful Dense Visual Affordance for Deformable Object Manipulation
Ruihai Wu (Peking University), Hao Dong (Peking University)
Robotic IntelligenceConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: A Foresightful Dense Visual Affordance framework is proposed, utilizing self-supervised multi-stage training and Fold-to-Unfold learning for point-to-point pick-and-place strategies applicable to multi-step flexible object manipulation.
Learning from Noisy Data for Semi-Supervised 3D Object Detection
Zehui Chen (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
Object DetectionAutonomous DrivingPoint Cloud
🎯 What it does: This paper proposes a pseudo-labeling based semi-supervised 3D object detection framework called NoiseDet, which effectively reduces the negative impact of pseudo-label noise on the model by softening classification targets, probability regression, and pixel-level feature consistency constraints.
Learning from Noisy Pseudo Labels for Semi-Supervised Temporal Action Localization
Kun Xia (Xi'an Jiaotong University), Wei Tang (University of Illinois Chicago)
RecognitionObject DetectionTransformerVideo
🎯 What it does: A noise pseudo-label learning framework for semi-supervised temporal action localization is proposed, which enhances the quality of pseudo-labels and suppresses localization bias through ranking, filtering, and consistency learning.
Learning from Semantic Alignment between Unpaired Multiviews for Egocentric Video Recognition
Qitong Wang (University of Delaware), Xi Peng (Google Research)
RecognitionLarge Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes a cross-view learning method SUM-L for unpaired, semantically partially similar front-view and back-view videos. It utilizes a large language model to extract video text semantics, constructs pseudo-pairs, and learns view-invariant features through semantic-aware contrastive learning and video-text multimodal alignment, ultimately improving action recognition performance in front-view videos.
Learning Gabor Texture Features for Fine-Grained Recognition
Lanyun Zhu (Singapore University of Technology and Design), Jun Liu (Singapore University of Technology and Design)
RecognitionConvolutional Neural NetworkImage
🎯 What it does: A fine-grained recognition framework is proposed, combining the CNN semantic branch with a texture branch based on Gabor filters, utilizing learnable Gabor filters equipped with statistical feature extractors and region selection gates to achieve efficient extraction and fusion of texture information.
Learning Global-aware Kernel for Image Harmonization
Xintian Shen (Zhejiang University), Yong Liu (Shanghai Jiao Tong University)
Image HarmonizationRestorationTransformerImage
🎯 What it does: A Global Perception Kernel Network (GKNet) is proposed for image harmonization, combining global long-distance background information with local details to achieve local-global dual compensation.
Learning Hierarchical Features with Joint Latent Space Energy-Based Prior
Jiali Cui (Stevens Institute of Technology), Tian Han (University of California)
GenerationAnomaly DetectionAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: This study investigates how to use energy-based prior with a joint latent space in multi-layer generative models to learn hierarchical features and proposes a joint training framework.
Learning Human Dynamics in Autonomous Driving Scenarios
Jingbo Wang (NVIDIA), Sameh Khamis (NVIDIA)
Autonomous DrivingReinforcement LearningPoint Cloud
🎯 What it does: In autonomous driving scenarios, using physical constraints and reinforcement learning controllers, we learn and generate human motion sequences that maintain physical feasibility even when obscured by vehicles or pedestrians under a moving camera;
Learning Human-Human Interactions in Images from Weak Textual Supervision
Morris Alper (Tel Aviv University), Hadar Averbuch-Elor (Tel Aviv University)
GenerationKnowledge DistillationTransformerLarge Language ModelVision Language ModelImageText
🎯 What it does: Learning free-text descriptions of human-human interaction (HHI) in a single image through weak text supervision;
Learning Image Harmonization in the Linear Color Space
Ke Xu (City University of Hong Kong), Rynson W.H. Lau (City University of Hong Kong)
Image HarmonizationRestorationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: This paper proposes an image harmonization method based on linear color space. First, a diffusion-based image deprocessing module is used to restore the composite image to a high dynamic range linear image. Then, the foreground colors are harmonized in this linear space and remapped back to sRGB, ultimately resulting in a visually more natural composite image.
Learning Image-Adaptive Codebooks for Class-Agnostic Image Restoration
Kechun Liu (University of Washington), Jinwei Gu (Chinese University of Hong Kong)
RestorationSuper ResolutionTransformerAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: An adaptive codebook called AdaCode is proposed for arbitrary image reconstruction, super-resolution, and inpainting.
Learning in Imperfect Environment: Multi-Label Classification with Long-Tailed Distribution and Partial Labels
Wenqiao Zhang (Zhejiang University), Yueting Zhuang (Zhejiang University)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a multi-label classification task with the coexistence of long-tail distribution and partial labels (PLT-MLC), and designs an end-to-end COMIC framework for learning.
Learning Long-Range Information with Dual-Scale Transformers for Indoor Scene Completion
Ziqi Wang (Wuhan University), Chunxia Xiao (Wuhan University)
RestorationTransformerContrastive LearningPoint Cloud
🎯 What it does: This paper proposes a Dual-Scale Transformer Network (DST-Net) and a two-stage completion strategy, first completing at the block level and then refining at the full scene level, aimed at addressing geometric missing data in indoor 3D environments.
Learning Navigational Visual Representations with Semantic Map Supervision
Yicong Hong (Adobe Research), Hao Tan (Adobe Research)
Representation LearningRobotic IntelligenceTransformerContrastive LearningImage
🎯 What it does: A navigation visual representation learning method based on contrastive learning, called Ego-Map, is proposed. It aligns the ego-centric images with the semantic maps generated from their trajectories, thereby enhancing the ability to capture the semantic and spatial information required for navigation.
Learning Neural Eigenfunctions for Unsupervised Semantic Segmentation
Zhijie Deng (Shanghai Jiao Tong University), Yucen Luo (Max Planck Institute for Intelligent Systems)
SegmentationTransformerImage
🎯 What it does: This study proposes a neural network-based spectral clustering method that learns spectral embeddings for unsupervised semantic segmentation using neural feature functions and directly outputs discrete clustering results.
Learning Neural Implicit Surfaces with Object-Aware Radiance Fields
Yiheng Zhang, Tao Mei
SegmentationGenerationNeural Radiance FieldPoint CloudMeshBenchmark
🎯 What it does: The study of neural implicit surface models based on volumetric rendering can directly reconstruct the geometry of foreground objects without the need for manual masks.
Learning Non-Local Spatial-Angular Correlation for Light Field Image Super-Resolution
Zhengyu Liang (National University of Defense Technology), Yulan Guo (National University of Defense Technology)
RestorationSuper ResolutionTransformerImage
🎯 What it does: To address the super-resolution problem of light field images, this paper proposes a Transformer-based network (EPIT) that learns non-local spatial-angular correlations by rearranging the 4D light field into multiple 2D EPIs (epipolar plane images), thereby achieving high-quality high-resolution light field reconstruction.
Learning Optical Flow from Event Camera with Rendered Dataset
Xinglong Luo (University of Electronic Science and Technology of China), Shuaicheng Liu (Hong Kong University of Science and Technology)
Data SynthesisOptical FlowImageVideo
🎯 What it does: A new event optical flow dataset MDR based on graphic rendering is proposed, along with the design of an Adaptive Density Module (ADM) to enhance the robustness of event optical flow estimation.
Learning Point Cloud Completion without Complete Point Clouds: A Pose-Aware Approach
Jihun Kim (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
RestorationGenerationPose EstimationComputational EfficiencyAuto EncoderPoint Cloud
🎯 What it does: A point cloud completion framework is proposed that does not require complete point clouds during the training phase. It generates multi-view incomplete clouds from a single incomplete point cloud using a pose-aware projection method, and then integrates these incomplete clouds to obtain a complete point cloud.
Learning Pseudo-Relations for Cross-domain Semantic Segmentation
Dong Zhao (Xidian University), Licheng Jiao (Xidian University)
SegmentationDomain AdaptationKnowledge DistillationImage
🎯 What it does: A cross-domain semantic segmentation framework RTea based on pseudo-relation learning is proposed, which utilizes a pixel relationship matrix generated by pseudo-labels to guide model learning.
Learning Rain Location Prior for Nighttime Deraining
Fan Zhang (Beijing Institute of Technology), Ying Fu (Beijing Institute of Technology)
RestorationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper studies the problem of raindrop removal at night, proposing to learn the Rain Location Prior (RLP) through a recursive residual network, and utilizing the Rain Prior Injection Module (RPIM) with an attention mechanism to inject this prior into U-shaped networks or Transformers to enhance rain removal performance.
Learning Robust Representations with Information Bottleneck and Memory Network for RGB-D-based Gesture Recognition
Yunan Li (Xidian University), Qiguang Miao (Xidian University)
RecognitionRepresentation LearningConvolutional Neural NetworkContrastive LearningImageMultimodality
🎯 What it does: Robust RGB-D gesture recognition feature representation is learned through information bottleneck theory and memory networks.
Learning Semi-supervised Gaussian Mixture Models for Generalized Category Discovery
Bingchen Zhao (University of Edinburgh), Kai Han (University of Hong Kong)
ClassificationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: This paper proposes an EM-like framework that dynamically estimates the number of categories and prototypes using a semi-supervised Gaussian mixture model, and completes clustering of unlabeled data through prototype contrastive learning, achieving universal category discovery without prior knowledge of the number of categories.
Learning Shape Primitives via Implicit Convexity Regularization
Xiaoyang Huang (Shanghai Jiao Tong University), Bingbing Ni (Shanghai Jiao Tong University)
SegmentationGenerationImagePoint Cloud
🎯 What it does: This paper proposes a new implicit convexity regularization (ICR) that achieves shape primitive decomposition without 3D data by learning implicit shape primitives from multi-view images.
Learning Spatial-context-aware Global Visual Feature Representation for Instance Image Retrieval
Zhongyan Zhang (University of Wollongong), Piotr Koniusz (CSIRO)
RetrievalConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: An end-to-end feature learning framework is proposed, which constructs visual feature branches and spatial context branches in parallel on a CNN backbone. It captures local descriptor types and their spatial distribution through online token learning and random walk-based distance encoding, and fuses them with visual features using cross-attention to generate spatial context-aware global features for instance image retrieval.
Learning Support and Trivial Prototypes for Interpretable Image Classification
Chong Wang (Australian Institute for Machine Learning), Gustavo Carneiro (University of Surrey)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: Proposes ST-ProtoPNet, which integrates support (near the classification boundary) prototypes with traditional trivial prototypes to achieve interpretable image classification.
Learning Symmetry-Aware Geometry Correspondences for 6D Object Pose Estimation
Heng Zhao (Hikvision Research Institute), Shiliang Pu (Hikvision Research Institute)
Pose EstimationConvolutional Neural NetworkPoint Cloud
🎯 What it does: This paper proposes a 6D object pose estimation framework called GCPose based on geometric correspondence, which can estimate the pose of any unseen object without the need for retraining.
Learning to Distill Global Representation for Sparse-View CT
Zilong Li (Fudan University), Hongming Shan (Fudan University)
RestorationKnowledge DistillationContrastive LearningImageComputed Tomography
🎯 What it does: A sparse-angle CT reconstruction method based on global representation distillation, GloReDi, has been designed.
Learning to Generate Semantic Layouts for Higher Text-Image Correspondence in Text-to-Image Synthesis
Minho Park (Korea Advanced Institute of Science and Technology), Jaegul Choo (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisDiffusion modelImageTextMultimodality
🎯 What it does: Proposes a Gaussian-categorical sampling diffusion model that jointly generates images and semantic layouts to enhance text-image correspondence.
Learning to Ground Instructional Articles in Videos through Narrations
Effrosyni Mavroudi (Meta AI), Lorenzo Torresani (Meta AI)
TransformerContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes an unsupervised learning framework that achieves temporal localization of steps in instructional videos by jointly aligning video, narration, and wikiHow step information.
Learning to Identify Critical States for Reinforcement Learning from Videos
Haozhe Liu (King Abdullah University of Science and Technology), Jürgen Schmidhuber (King Abdullah University of Science and Technology)
Convolutional Neural NetworkRecurrent Neural NetworkReinforcement LearningVideo
🎯 What it does: A video-based deep state recognition framework called Deep State Identifier is proposed to automatically identify the most critical states for final rewards in the reinforcement learning process without action information.
Learning to Learn: How to Continuously Teach Humans and Machines
Parantak Singh (Nanyang Technological University), Mengmi Zhang (Nanyang Technological University)
OptimizationMeta LearningConvolutional Neural NetworkImage
🎯 What it does: This study investigates the impact of curriculum order on the performance of both humans and machine learning in online class-incremental continual learning, and proposes an automatic curriculum designer.
Learning to Transform for Generalizable Instance-wise Invariance
Utkarsh Singhal (University of California Berkeley), Stella X. Yu (University of Michigan)
ClassificationData-Centric LearningFlow-based ModelImage
🎯 What it does: This paper achieves flexible, adaptable, and transferable invariance to geometric transformations during testing by learning the distribution of input condition transformations.
Learning to Upsample by Learning to Sample
Wenze Liu (Huazhong University of Science and Technology), Zhiguo Cao (Huazhong University of Science and Technology)
Object DetectionSegmentationDepth EstimationImage
🎯 What it does: A lightweight and efficient dynamic upsampler called DySample is proposed, which utilizes point sampling to achieve feature upsampling, eliminating the need for traditional dynamic convolution, additional high-resolution guided features, and custom CUDA code.
Learning Trajectory-Word Alignments for Video-Language Tasks
Xu Yang (Southeast University), Songfang Huang (Alibaba Group)
RetrievalTransformerVision Language ModelVideoTextMultimodality
🎯 What it does: The TW-BERT model is proposed, which achieves trajectory-word alignment in video-language tasks through trajectory-word (T2W) attention and an asynchronous cross-encoder, and introduces a Hierarchical Frame Selection (HFS) module during the fine-tuning phase to filter the most relevant frames.
Learning Unified Decompositional and Compositional NeRF for Editable Novel View Synthesis
Yuxin Wang (Hong Kong University of Science and Technology), Dan Xu (Hong Kong University of Science and Technology)
GenerationData SynthesisNeural Radiance FieldImage
🎯 What it does: A unified NeRF framework is proposed, capable of simultaneously decomposing scenes (object-level and background radiance fields) and composing them, thus achieving editable panoramic synthesis and visualization.
Learning Versatile 3D Shape Generation with Improved Auto-regressive Models
Simian Luo (Fudan University), Xiangyang Xue (Fudan University)
GenerationData SynthesisTransformerGenerative Adversarial NetworkContrastive LearningImageTextPoint Cloud
🎯 What it does: An improved autoregressive model, ImAM, is proposed, which utilizes three-plane projection and coupling networks for discrete representation on latent vectors, achieving efficient 3D shape generation.
Learning Vision-and-Language Navigation from YouTube Videos
Kunyang Lin (South China University of Technology), Chuang Gan (UMass Amherst)
RecognitionRetrievalRobotic IntelligenceTransformerVision Language ModelVideoMultimodality
🎯 What it does: This paper constructs a large-scale VLN dataset (YouTube-VLN) using YouTube house tour videos and pre-trains the 'Lily' agent on this dataset to enable it to learn visual language navigation from natural videos.
Learning with Diversity: Self-Expanded Equalization for Better Generalized Deep Metric Learning
Jiexi Yan (Xidian University), Heng Huang (University of Maryland)
RetrievalDomain AdaptationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: An adaptive expansion and equalization (Self-Expanded Equalization, SEE) framework is proposed for single-source domain general deep metric learning.
Lecture Presentations Multimodal Dataset: Towards Understanding Multimodality in Educational Videos
Dong Won Lee (Carnegie Mellon University), Louis-Philippe Morency (Carnegie Mellon University)
GenerationRetrievalTransformerLarge Language ModelVision Language ModelImageVideoTextMultimodality
🎯 What it does: A large-scale Lecture Presentations Multimodal (LPM) dataset is proposed, along with three retrieval/generation tasks designed to evaluate visual language models' cross-modal understanding of educational videos.
Lens Parameter Estimation for Realistic Depth of Field Modeling
Dominique Piché-Meunier (Laval University), Jean-François Lalonde (Adobe)
Data SynthesisDepth EstimationTransformerSupervised Fine-TuningImage
🎯 What it does: Estimate the depth of field parameters (focal length difference and blur factor) of the camera lens from a single image, and use these parameters to achieve globally consistent depth modeling and 3D/2D object synthesis.
LERF: Language Embedded Radiance Fields
Justin Kerr (University of California Berkeley), Matthew Tancik (University of California Berkeley)
Object DetectionRetrievalTransformerNeural Radiance FieldContrastive LearningImage
🎯 What it does: Directly dense encoding of CLIP's image-text embeddings into NeRF to construct a multi-scale language field that allows zero-shot text queries in three-dimensional space.
Less is More: Focus Attention for Efficient DETR
Dehua Zheng (Huazhong University of Science and Technology), Yunhe Wang (Huawei Noah's Ark Lab)
Object DetectionComputational EfficiencyTransformerImage
🎯 What it does: A new object detection model called Focus-DETR is proposed, aimed at improving computational efficiency and model accuracy by focusing on more informative tokens.
Leveraging Inpainting for Single-Image Shadow Removal
Xiaoguang Li (University of South Carolina), Song Wang (University of South Carolina)
Image TranslationRestorationConvolutional Neural NetworkSupervised Fine-TuningGenerative Adversarial NetworkImage
🎯 What it does: Pre-trained image inpainting network, fine-tuned for single image shadow removal, and proposed a dual-encoder adaptive fusion network.
Leveraging Intrinsic Properties for Non-Rigid Garment Alignment
Siyou Lin (Tsinghua University), Yebin Liu (Tsinghua University)
Point CloudMeshBenchmark
🎯 What it does: A two-stage pipeline based on intrinsic manifold properties and neural deformation fields is used to achieve high-precision texture-level and geometric detail alignment for real 3D garment scans.
Leveraging SE(3) Equivariance for Learning 3D Geometric Shape Assembly
Ruihai Wu (Peking University), Hao Dong (Peking University)
Pose EstimationRobotic IntelligenceGraph Neural NetworkGenerative Adversarial NetworkPoint CloudMesh
🎯 What it does: Utilizing SE(3) equivariant networks to achieve geometric shape assembly tasks, that is, recovering complete objects from broken parts without semantic information.
Leveraging Spatio-Temporal Dependency for Skeleton-Based Action Recognition
Jungho Lee (Yonsei University), Sangyoun Lee (Korea Institute of Science and Technology)
RecognitionPose EstimationGraph Neural NetworkVideoMultimodality
🎯 What it does: A new skeleton action recognition framework called STC-Net is proposed, which enhances the model's spatiotemporal perception ability using the Spatiotemporal Curve (STC) module and the Dilated Graph Convolution (DK-GC) module.
LexLIP: Lexicon-Bottlenecked Language-Image Pre-Training for Large-Scale Image-Text Sparse Retrieval
Ziyang Luo (Hong Kong Baptist University), Daxin Jiang (Microsoft Corporation)
RetrievalTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: This paper proposes a sparse retrieval paradigm that maps images and text into a discrete dictionary space, using dictionary sparse vectors to achieve large-scale image-text retrieval.
LFS-GAN: Lifelong Few-Shot Image Generation
Juwon Seo (Kyung Hee University), Gyeong-Moon Park (Kyung Hee University)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: Proposes the LFS-GAN framework to achieve lifelong few-shot image generation, addressing the issues of catastrophic forgetting and mode collapse.
LiDAR-Camera Panoptic Segmentation via Geometry-Consistent and Semantic-Aware Alignment
Zhiwei Zhang (Shanghai Jiaotong University), Lizhuang Ma (Shanghai Jiaotong University)
SegmentationAutonomous DrivingMultimodalityPoint Cloud
🎯 What it does: A multi-modal LiDAR-camera fusion network LCPS is proposed for 3D panoramic segmentation, integrating four modules: ACPA, SARA, PVP, and FOG;
LiDAR-UDA: Self-ensembling Through Time for Unsupervised LiDAR Domain Adaptation
Amirreza Shaban (University of Washington), Byron Boots (University of Washington)
SegmentationDomain AdaptationAutonomous DrivingPoint Cloud
🎯 What it does: This paper proposes an unsupervised domain adaptation framework named LiDAR-UDA, which utilizes self-supervised temporal consistency to improve the transfer performance of LiDAR semantic segmentation models.
LightDepth: Single-View Depth Self-Supervision from Illumination Decline
Javier Rodríguez-Puigvert (I3A - Universidad de Zaragoza), Javier Civera (I3A - Universidad de Zaragoza)
Depth EstimationConvolutional Neural NetworkImage
🎯 What it does: A single-view self-supervised depth estimation method called LightDepth is proposed, which utilizes light attenuation for self-supervised training in endoscopic images.
LightGlue: Local Feature Matching at Light Speed
Philipp Lindenberger (ETH Zurich), Marc Pollefeys (Microsoft Mixed Reality and AI Lab)
Pose EstimationTransformerImage
🎯 What it does: LightGlue is proposed, a lightweight deep matching network based on Transformer, designed for fast and accurate matching of sparse local features between image pairs.
Lighting Every Darkness in Two Pairs: A Calibration-Free Pipeline for RAW Denoising
Xin Jin (Nankai University), Chongyi Li (Nankai University)
RestorationImage
🎯 What it does: A calibration-free raw image denoising pipeline LED is proposed, capable of directly denoising under different cameras and gains.
Lighting up NeRF via Unsupervised Decomposition and Enhancement
Haoyuan Wang (City University of Hong Kong), Rynson W.H. Lau (City University of Hong Kong)
RestorationGenerationNeural Radiance FieldImage
🎯 What it does: For low-light scene sRGB images, an unsupervised method is proposed to perform illumination and color decomposition in the implicit scene representation of NeRF, directly learning to enhance brightness, denoise, and color correct, thereby generating new perspective images with high visibility, realistic colors, and rich details.
Lightweight Image Super-Resolution with Superpixel Token Interaction
Aiping Zhang (Sun Yat-Sen University), Xiaochun Cao (Sun Yat-Sen University)
RestorationSuper ResolutionConvolutional Neural NetworkTransformerImage
🎯 What it does: A lightweight single-image super-resolution network SPIN is proposed, utilizing superpixel partitioning and Transformer, which integrates superpixel aggregation, cross-superpixel attention, and internal superpixel attention;
LIMITR: Leveraging Local Information for Medical Image-Text Representation
Gefen Dawidowicz (Technion Israel Institute of Technology), Ayellet Tal (Technion Israel Institute of Technology)
RetrievalRepresentation LearningTransformerContrastive LearningImageTextBiomedical Data
🎯 What it does: This paper proposes a model named LIMITR, which is used to learn the joint representation of chest X-ray images and their corresponding radiology reports, and applies it to tasks such as text-image retrieval, phrase localization, and category retrieval.
Linear Spaces of Meanings: Compositional Structures in Vision-Language Models
Matthew Trager (Amazon Web Services Artificial Intelligence Labs), Stefano Soatto (Amazon Web Services Artificial Intelligence Labs)
ClassificationRetrievalTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: The paper studies the linear decomposable structure of visual-language model embeddings and introduces the concept of 'ideal words', explaining and utilizing these structures to complete tasks such as classification, debiasing, and retrieval through simple linear operations.
Linear-Covariance Loss for End-to-End Learning of 6D Pose Estimation
Fulin Liu (Beihang University), Mathieu Salzmann (EPFL)
Pose EstimationImage
🎯 What it does: This paper proposes a loss function based on linear covariance for end-to-end learning of 6D pose estimation.
LinkGAN: Linking GAN Latents to Pixels for Controllable Image Synthesis
Jiapeng Zhu (Hong Kong University of Science and Technology), Qifeng Chen (Hong Kong University of Science and Technology)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: A regularization term is added during GAN training to establish explicit associations between certain latent space axes and specific pixels or regions in the image, enabling locally controllable image synthesis.
Lip Reading for Low-resource Languages by Learning and Combining General Speech Knowledge and Language-specific Knowledge
Minsu Kim (Korea Advanced Institute of Science and Technology), Yong Man Ro (Korea Advanced Institute of Science and Technology)
RecognitionTransformerSupervised Fine-TuningVideoTextAudio
🎯 What it does: A lip-reading framework for low-resource languages is proposed, which first learns general speech knowledge through masked prediction on a high-resource language (English), then trains a language-specific memory-enhanced decoder (LMDecoder) on audio-text paired data of the target language, and finally couples the two modules through an attention mechanism to achieve low-resource lip-reading.
Lip2Vec: Efficient and Robust Visual Speech Recognition via Latent-to-Latent Visual to Audio Representation Mapping
Yasser Abdelaziz Dahou Djilali (Technology Innovation Institute), Merouane Debbah (Technology Innovation Institute)
RecognitionRepresentation LearningTransformerContrastive LearningVideoAudio
🎯 What it does: A framework called Lip2Vec is proposed to reconstruct visual speech recognition into audio speech recognition, achieving text output using only video by learning the mapping from lip features to audio latent space.
LIST: Learning Implicitly from Spatial Transformers for Single-View 3D Reconstruction
Mohammad Samiul Arshad (University of Texas at Arlington), William J. Beksi (University of Texas at Arlington)
GenerationDepth EstimationConvolutional Neural NetworkImagePoint Cloud
🎯 What it does: This paper proposes the LIST network, which utilizes both local and global features and achieves implicit reconstruction of 3D objects from a single view through spatial transformation, avoiding the need for camera estimation and pixel alignment.
LISTER: Neighbor Decoding for Length-Insensitive Scene Text Recognition
Changxu Cheng (DAMO Academy), Cong Yao (DAMO Academy)
RecognitionTransformerImageText
🎯 What it does: A length-insensitive scene text recognition framework called LISTER is proposed, which maintains high accuracy on both short and long texts.
LiveHand: Real-time and Photorealistic Neural Hand Rendering
Akshay Mundra (Max Planck Institute for Informatics), Mohamed Elgharib (Max Planck Institute for Informatics)
GenerationPose EstimationNeural Radiance FieldMesh
🎯 What it does: This paper presents LiveHand, a real-time, illumination-variable hand rendering method based on neural implicit fields.
LivelySpeaker: Towards Semantic-Aware Co-Speech Gesture Generation
Yihao Zhi (ShanghaiTech University), Shenghua Gao (ShanghaiTech University)
GenerationTransformerDiffusion modelVideoAudio
🎯 What it does: This paper presents the LivelySpeaker system, which splits the generation of speech-accompanied gestures into two stages: semantic-aware generation and rhythm-aware generation. These stages are implemented using CLIP text embeddings and an MLP-based diffusion model, ultimately achieving gesture animations that are both semantically and rhythmically aligned.
LivePose: Online 3D Reconstruction from Monocular Video with Dynamic Camera Poses
Noah Stier (Apple), Ming Chuang (University of California)
Pose EstimationDepth EstimationConvolutional Neural NetworkRecurrent Neural NetworkTransformerSimultaneous Localization and MappingVideoPoint Cloud
🎯 What it does: The paper proposes a 3D reconstruction framework capable of online processing of dynamic camera poses, addressing the issue of geometric inconsistency caused by traditional methods neglecting pose updates.