ICCV 2023 Papers — Page 22
IEEE/CVF International Conference on Computer Vision · 2156 papers
VoroMesh: Learning Watertight Surface Meshes with Voronoi Diagrams
Nissim Maruani (Inria, University Côte d'Azur), Mathieu Desbrun (Inria)
GenerationOptimizationConvolutional Neural NetworkMesh
🎯 What it does: A differentiable Voronoi-based mesh representation called VoroMesh is proposed, along with the design of VoroLoss, which is used to directly optimize the positions and occupancy of generators predicted by neural networks, resulting in closed and non-self-intersecting 3D surfaces.
Vox-E: Text-Guided Voxel Editing of 3D Objects
Etai Sella (Tel Aviv University), Hadar Averbuch-Elor (Tel Aviv University)
GenerationData SynthesisDiffusion modelScore-based ModelMesh
🎯 What it does: Construct a voxel grid using multi-view images and utilize a pre-trained diffusion model for text-driven geometric and appearance editing.
VQ3D: Learning a 3D-Aware Generative Model on ImageNet
Kyle Sargent (Stanford University), Deqing Sun (Google Research)
GenerationData SynthesisTransformerNeural Radiance FieldAuto EncoderImage
🎯 What it does: A two-stage vector quantization Autoencoder has been designed and trained, where Stage 1 uses a variable plane (triplane) + NeRF decoder to achieve 3D-aware image reconstruction and viewpoint transformation, and Stage 2 employs a Transformer for autoregressive generation of discrete codes, capable of generating 3D structurally consistent images on large-scale multi-class datasets like ImageNet.
VQA Therapy: Exploring Answer Differences by Visually Grounding Answers
Chongyan Chen (University of Texas at Austin), Danna Gurari (University of Colorado Boulder)
TransformerVision Language ModelImageText
🎯 What it does: A VQA-AnswerTherapy dataset was constructed, performing visual grounding for all valid answers to each visual question, and proposing two new tasks: single and multiple answer grounding prediction, and answer grounding localization.
VQA-GNN: Reasoning with Multimodal Knowledge via Graph Neural Networks for Visual Question Answering
Yanan Wang (KDDI Research), Jure Leskovec (Stanford University)
RecognitionRetrievalGraph Neural NetworkVision Language ModelImageMultimodality
🎯 What it does: A VQA-GNN model is designed to achieve bidirectional fusion and reasoning of structured and unstructured knowledge by constructing a multimodal semantic graph of scene graphs, concept graphs, and QA context.
Waffling Around for Performance: Visual Classification with Random Words and Broad Concepts
Karsten Roth (University of Tuebingen), Zeynep Akata (Google DeepMind)
ClassificationLarge Language ModelVision Language ModelImage
🎯 What it does: A zero-shot image classification method called Waffle CLIP is proposed, which replaces the fine-grained descriptions originally generated by LLMs with random words or character sequences added to the category prompts; it also proposes using LLMs to automatically extract high-level concepts to alleviate category ambiguity.
WALDO: Future Video Synthesis Using Object Layer Decomposition and Parametric Flow Prediction
Guillaume Le Moing (Inria), Cordelia Schmid (Inria)
SegmentationGenerationData SynthesisPose EstimationTransformerOptical FlowVideo
🎯 What it does: WALDO predicts future frames by automatically decomposing each frame of video into multiple object layers and using control points and thin plate splines for motion modeling.
Walking Your LiDOG: A Journey Through Multiple Domains for LiDAR Semantic Segmentation
Cristiano Saltori (University of Trento), Laura Leal-Taixé (NVIDIA)
SegmentationDomain AdaptationAutonomous DrivingConvolutional Neural NetworkContrastive LearningPoint Cloud
🎯 What it does: This paper studies the domain generalization problem of LiDAR semantic segmentation under different sensors and environments, proposing and validating a novel LiDOG framework.
Wasserstein Expansible Variational Autoencoder for Discriminative and Generative Continual Learning
Fei Ye (University of York), Adrian G. Bors (University of York)
ClassificationGenerationData SynthesisAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: A dynamically expandable Wasserstein Variational Autoencoder (WEVAE) is proposed, which determines when to add new components using Wasserstein distance and filters memory samples with an energy function to enhance knowledge diversity.
WaterMask: Instance Segmentation for Underwater Imagery
Shijie Lian (Hainan University), Sam Kwong (City University of Hong Kong)
Object DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: The first underwater instance segmentation dataset UIIS is proposed, along with the development of a dedicated model WaterMask.
WaveIPT: Joint Attention and Flow Alignment in the Wavelet domain for Pose Transfer
Liyuan Ma (Alibaba Group), Kejie Huang (Zhejiang University)
Image TranslationGenerationPose EstimationConvolutional Neural NetworkGenerative Adversarial NetworkOptical FlowImage
🎯 What it does: The WaveIPT method is proposed, which integrates attention and optical flow in the waveform domain to achieve more refined human pose transfer image generation.
WaveNeRF: Wavelet-based Generalizable Neural Radiance Fields
Muyu Xu (Nanyang Technological University), Shijian Lu (Nanyang Technological University)
GenerationData SynthesisNeural Radiance FieldAuto EncoderImage
🎯 What it does: This paper proposes WaveNeRF, which combines wavelet frequency domain multi-view stereo with NeRF to achieve high-quality generalization of novel view synthesis without scene fine-tuning.
WDiscOOD: Out-of-Distribution Detection via Whitened Linear Discriminant Analysis
Yiye Chen (Georgia Institute of Technology), Patricio A. Vela (Georgia Institute of Technology)
Anomaly DetectionTransformerContrastive LearningImage
🎯 What it does: A method for OOD detection based on Whitened Linear Discriminant Analysis (WLDA) called WDiscOOD is designed, which projects features into the discriminative subspace and the residual subspace, and generates OOV scores through a weighted combination of the distances between the two.
Weakly Supervised Learning of Semantic Correspondence through Cascaded Online Correspondence Refinement
Yiwen Huang (Fudan University), Weifeng Ge (Fudan University)
RecognitionSegmentationTransformerImage
🎯 What it does: A weakly supervised semantic correspondence framework based on multi-instance learning is proposed, which can learn pixel-level correspondence using only image-level labels.
Weakly Supervised Referring Image Segmentation with Intra-Chunk and Inter-Chunk Consistency
Jungbeom Lee (Amazon), Tara Taghavi (Amazon)
Object DetectionSegmentationTransformerVision Language ModelImageText
🎯 What it does: This paper proposes a weakly supervised referential image segmentation method that only utilizes image-text pairs, achieving high-quality segmentation results without pixel-level annotations.
Weakly-supervised 3D Pose Transfer with Keypoints
Jinnan Chen (National University of Singapore), Gim Hee Lee (National University of Singapore)
Pose EstimationMesh
🎯 What it does: A weakly supervised 3D pose transfer framework is proposed, which achieves pose transfer for different topological meshes with only keypoint supervision while keeping the source shape unchanged.
Weakly-Supervised Action Localization by Hierarchically-Structured Latent Attention Modeling
Guiqin Wang (Xi'an Jiao Tong University), Qinghai Guo (Huawei Technologies)
RecognitionObject DetectionRecurrent Neural NetworkAuto EncoderVideo
🎯 What it does: A weakly supervised action localization framework AHLM is proposed, which utilizes a hierarchical latent attention model to simultaneously detect feature semantic change points and accurately locate action boundaries.
Weakly-Supervised Action Segmentation and Unseen Error Detection in Anomalous Instructional Videos
Reza Ghoddoosian (Honda Research Institute), Behzad Dariush (Honda Research Institute)
SegmentationAnomaly DetectionTransformerOptical FlowVideo
🎯 What it does: A weakly supervised action segmentation and unseen error detection method is proposed, utilizing an unconstrained Viterbi algorithm to segment abnormal teaching videos and detect errors in real-time conditions.
Weakly-Supervised Text-Driven Contrastive Learning for Facial Behavior Understanding
Xiang Zhang (State University of New York at Binghamton), Lijun Yin (State University of New York at Binghamton)
RecognitionRepresentation LearningContrastive LearningImageTextMultimodality
🎯 What it does: A weakly supervised text-driven contrastive learning framework called CLEF is proposed, which enhances the representation and recognition performance of facial expressions and action units using activity descriptions and label texts.
What Can a Cook in Italy Teach a Mechanic in India? Action Recognition Generalisation Over Scenarios and Locations
Chiara Plizzari (Politecnico di Torino), Dima Damen (University of Bristol)
RecognitionDomain AdaptationContrastive LearningVideoText
🎯 What it does: The authors propose a new problem of action recognition generalization and construct a large-scale ARGO1M dataset to evaluate action recognition in unseen scenes and locations.
What can Discriminator do? Towards Box-free Ownership Verification of Generative Adversarial Networks
Ziheng Huang (Wuhan University), Lina Wang (Nanjing University of Aeronautics and Astronautics)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: A box-free GAN ownership verification method is designed, which proves copyright solely by detecting generated images using the unique distribution learned by the generator.
What Can Simple Arithmetic Operations Do for Temporal Modeling?
Wenhao Wu (University of Sydney), Wanli Ouyang (Shanghai AI Laboratory)
ClassificationRecognitionOptimizationConvolutional Neural NetworkTransformerVideo
🎯 What it does: By using the four simplest arithmetic operations of addition, subtraction, multiplication, and division between frame features to generate auxiliary temporal information and embedding it into the original features, temporal modeling of videos is achieved.
What do neural networks learn in image classification? A frequency shortcut perspective
Shunxin Wang (University of Twente), Nicola Strisciuglio (University of Twente)
ClassificationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper systematically analyzes the learning dynamics of neural networks in image classification tasks from a frequency perspective. It finds that networks tend to utilize specific frequency subsets (frequency shortcuts) to quickly complete classification. It proposes a frequency feature metric based on Fourier transform (ADCS) and a frequency shortcut identification method (DFM), and further evaluates the performance and transferability of model capacity, data augmentation, and frequency shortcuts in both ID and OOD environments.
What Does a Platypus Look Like? Generating Customized Prompts for Zero-Shot Image Classification
Sarah Pratt (University of Washington), Ali Farhadi (University of Washington)
ClassificationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImage
🎯 What it does: Proposes the CuPL method, which uses LLM to automatically generate customized prompts for zero-shot image classification.
What does CLIP know about a red circle? Visual prompt engineering for VLMs
Aleksandar Shtedritski (University of Oxford), Andrea Vedaldi (University of Oxford)
ClassificationObject DetectionRetrievalTransformerPrompt EngineeringVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes a visual prompting technique (marking) by drawing red circles on images, utilizing the zero-shot capability of CLIP to accomplish traditional visual tasks such as reference expression understanding and keypoint localization, which are difficult to achieve solely through cropping.
When Do Curricula Work in Federated Learning?
Saeed Vahidian (University of California San Diego), Bill Lin (University of California San Diego)
Federated LearningImage
🎯 What it does: This paper studies how Curriculum Learning (CL) can alleviate the problem of client data heterogeneity in federated learning and explores the impact of the ordered learning principle on federated learning.
When Epipolar Constraint Meets Non-Local Operators in Multi-View Stereo
Tianqi Liu (Huazhong University of Science and Technology), Zhiguo Cao (Huazhong University of Science and Technology)
Depth EstimationOptimizationComputational EfficiencyTransformerImageBenchmark
🎯 What it does: This paper proposes a non-local feature aggregation method based on disparity geometric constraints—Epipolar Transformer—to improve feature matching quality and reconstruction accuracy in multi-view stereo reconstruction.
When Noisy Labels Meet Long Tail Dilemmas: A Representation Calibration Method
Manyi Zhang (Tsinghua University), Weiran Huang (Shanghai Jiao Tong University)
Anomaly DetectionRepresentation LearningContrastive LearningGaussian SplattingImage
🎯 What it does: A representation calibration method based on contrastive learning (RCAL) is proposed, which first obtains robust features using unsupervised contrastive learning, and then performs distributed calibration and individual calibration on the representation distribution of each category, thereby addressing both noisy labels and long-tail distributions simultaneously.
When Prompt-based Incremental Learning Does Not Meet Strong Pretraining
Yu-Ming Tang (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
ClassificationRecognitionTransformerPrompt EngineeringImage
🎯 What it does: This paper proposes a learnable Adaptive Prompt Generator (APG) that replaces traditional fixed prompt pools by generating task-specific prompts in incremental learning, thereby reducing the semantic gap between pre-training tasks and subsequent tasks, and supporting zero-shot incremental learning.
When to Learn What: Model-Adaptive Data Augmentation Curriculum
Chengkai Hou (Jilin University), Tianyi Zhou (University of Maryland)
OptimizationData-Centric LearningConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: An adaptive data augmentation method called MADAug is designed, which dynamically selects augmentation operations based on the model training phase and sample information through an online learning strategy network, forming an augmentation curriculum from easy to difficult.
Who Are You Referring To? Coreference Resolution In Image Narrations
Arushi Goel (University of Edinburgh), Hakan Bilen (University of Edinburgh)
TransformerImageTextMultimodality
🎯 What it does: This paper proposes the task of coreference resolution in image narratives and predicts coreference chains along with corresponding image regions using a weakly supervised multimodal approach.
Why do networks have inhibitory/negative connections?
Qingyang Wang (Johns Hopkins University), Joshua T. Vogelstein (Johns Hopkins University)
Convolutional Neural Network
🎯 What it does: Proves that deep networks with non-negative weights cannot be universal approximators, indicating that negative weights are a necessary condition for enhancing representational capacity.
Why Is Prompt Tuning for Vision-Language Models Robust to Noisy Labels?
Cheng-En Wu (University of Wisconsin-Madison), Linjie Yang (ByteDance Inc.)
ClassificationRecognitionTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: This paper studies how prompt tuning of visual-language models (such as CLIP) maintains high performance under noisy labels, demonstrating their inherent robustness to noise and further enhancing unsupervised prompt tuning through the use of Generalized Cross Entropy (GCE) loss and random pseudo-labeling.
Will Large-scale Generative Models Corrupt Future Datasets?
Ryuichiro Hataya (RIKEN), Hiromi Arai (RIKEN)
ClassificationGenerationData SynthesisConvolutional Neural NetworkDiffusion modelImageText
🎯 What it does: This study investigates the impact of images generated by large text-to-image generation models on the performance of future datasets and computer vision models, by simulating the replacement of some real images in ImageNet and COCO with images generated by Stable Diffusion for evaluation.
Window-Based Early-Exit Cascades for Uncertainty Estimation: When Deep Ensembles are More Efficient than Single Models
Guoxuan Xia (Imperial College London), Christos-Savvas Bouganis (Imperial College London)
ClassificationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: This paper studies the computational-performance trade-off of Deep Ensembles and single model scaling in uncertainty estimation tasks, and proposes an early exit cascade strategy for decision boundary windows to significantly reduce inference costs while ensuring the quality of uncertainty.
With a Little Help from Your Own Past: Prototypical Memory Networks for Image Captioning
Manuele Barraco (University of Modena and Reggio Emilia), Rita Cucchiara (Istituto Italiano di Tecnologia)
GenerationRetrievalTransformerImageTextMultimodality
🎯 What it does: A Prototypical Memory Attention (PMA) network is designed, utilizing the activations of past training samples as learnable memory vectors directly embedded in the self-attention layer of the Transformer for image caption generation.
Workie-Talkie: Accelerating Federated Learning by Overlapping Computing and Communications via Contrastive Regularization
Rui Chen (University of Houston), Miao Pan (University of Houston)
Federated LearningComputational EfficiencyContrastive LearningImage
🎯 What it does: Accelerating training by overlapping local computation and communication in federated learning and introducing contrastive regularization.
X-Mesh: Towards Fast and Accurate Text-driven 3D Stylization via Dynamic Textual Guidance
Yiwei Ma (Xiamen University), Rongrong Ji (Xiamen University)
GenerationData SynthesisContrastive LearningMeshBenchmark
🎯 What it does: The X-Mesh framework is proposed, achieving text-based 3D mesh stylization by predicting the colors and geometric properties of vertices to match the bare mesh with the given text description.
X-VoE: Measuring eXplanatory Violation of Expectation in Physical Events
Bo Dai (Peking University), Yixin Zhu (Peking University)
Explainability and InterpretabilityTransformerAuto EncoderVideoPhysics Related
🎯 What it does: The X-VoE explanatory violation experiment dataset and the XPL model are proposed to evaluate and enhance AI's understanding and explanatory capabilities of intuitive physics.
XiNet: Efficient Neural Networks for tinyML
Alberto Ancilotto (Fondazione Bruno Kessler), Elisabetta Farella (Fondazione Bruno Kessler)
Object DetectionComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A XiNet neural network architecture suitable for TinyML is proposed, emphasizing direct hardware efficiency metrics, standard convolution blocks, mixed attention, and broadcast skip connections.
XMem++: Production-level Video Segmentation From Few Annotated Frames
Maksym Bekuzarov, Hao Li
SegmentationConvolutional Neural NetworkVideo
🎯 What it does: XMem++ introduces a permanent memory module, achieving high-quality video segmentation with minimal annotations, supporting multi-frame annotations and real-time interaction.
XNet: Wavelet-Based Low and High Frequency Fusion Networks for Fully- and Semi-Supervised Semantic Segmentation of Biomedical Images
Yanfeng Zhou (University of Chinese Academy of Sciences), Ge Yang (University of Chinese Academy of Sciences)
SegmentationConvolutional Neural NetworkImageMultimodalityBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A low-frequency (LF) and high-frequency (HF) fusion network named XNet is proposed, which can be used for both fully supervised and semi-supervised medical image semantic segmentation.
XVO: Generalized Visual Odometry via Cross-Modal Self-Training
Lei Lai (Boston University), Eshed Ohn-Bar (Boston University)
Autonomous DrivingOptimizationTransformerSimultaneous Localization and MappingOptical FlowVideoMultimodality
🎯 What it does: We propose XVO, a calibration-free and generalizable monocular visual odometry method based on cross-modal self-supervised training.
Yes, we CANN: Constrained Approximate Nearest Neighbors for Local Feature-Based Visual Localization
Dror Aiger (Google Research), Simon Lynen (Google Research)
Object DetectionRetrievalSimultaneous Localization and MappingImage
🎯 What it does: The CANN algorithm is proposed, which enhances visual localization by simultaneously searching in appearance and geometric space using local features.
You Never Get a Second Chance To Make a Good First Impression: Seeding Active Learning for 3D Semantic Segmentation
Nermin Samet (Ecole des Ponts), Vincent Lepetit (Ecole des Ponts)
SegmentationConvolutional Neural NetworkContrastive LearningPoint Cloud
🎯 What it does: This paper proposes the SeedAL method, which automatically constructs initial annotation seeds for 3D point cloud semantic segmentation, significantly enhancing the effectiveness of active learning.
Your Diffusion Model is Secretly a Zero-Shot Classifier
Alexander C. Li (Carnegie Mellon University), Deepak Pathak (Carnegie Mellon University)
ClassificationRecognitionDiffusion modelImage
🎯 What it does: Proposes a Diffusion Classifier that utilizes the ELBO estimation of a pre-trained text-to-image diffusion model for zero-shot classification.
Zenseact Open Dataset: A Large-Scale and Diverse Multimodal Dataset for Autonomous Driving
Mina Alibeigi (Zenseact), Christoffer Petersson (Zenseact)
Object DetectionSegmentationAutonomous DrivingImageVideoMultimodalityPoint Cloud
🎯 What it does: The largest multimodal autonomous driving dataset in Europe, ZOD, has been released, containing high-resolution camera, LiDAR, and GNSS/IMU data, and providing multi-task annotations such as 2D/3D object detection, road instance/semantic segmentation, and traffic sign recognition, supporting long-distance perception and multi-task learning.
Zero-1-to-3: Zero-shot One Image to 3D Object
Ruoshi Liu (Columbia University), Carl Vondrick (Columbia University)
GenerationData SynthesisDiffusion modelImageMesh
🎯 What it does: We propose Zero-1-to-3, a zero-shot single-image novel viewpoint synthesis and 3D reconstruction method based on diffusion models, which can generate images from different camera viewpoints given only a single RGB image and further recover a 3D mesh.
Zero-guidance Segmentation Using Zero Segment Labels
Pitchaporn Rewatbowornwong (VISTEC), Supasorn Suwajanakorn (VISTEC)
SegmentationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: Proposes a zero-shot semantic segmentation task that uses pre-trained DINO and CLIP to automatically segment without a training set or text prompts, generating natural language labels for each region.
Zero-Shot Composed Image Retrieval with Textual Inversion
Alberto Baldrati (University of Florence), Alberto Del Bimbo (University of Florence)
RetrievalKnowledge DistillationVision Language ModelContrastive LearningImageText
🎯 What it does: A zero-shot combined image retrieval method called SEARLE is proposed, which maps reference images to pseudo-word tokens using text inversion, and retrieves them in the CLIP text space after concatenating with relative descriptions, completely independent of labeled data.
Zero-Shot Contrastive Loss for Text-Guided Diffusion Image Style Transfer
Serin Yang (Korea Advanced Institute of Science and Technology), Jong Chul Ye (Korea Advanced Institute of Science and Technology)
Image TranslationGenerationDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes a diffusion model method based on zero-shot contrastive loss (ZeCon) for text-guided image style transfer, which maintains content consistency without additional training or fine-tuning.
Zero-Shot Point Cloud Segmentation by Semantic-Visual Aware Synthesis
Yuwei Yang (Sichuan University), Yinjie Lei (Sichuan University)
SegmentationGenerative Adversarial NetworkContrastive LearningPoint Cloud
🎯 What it does: A zero-shot point cloud semantic segmentation method based on semantic-visual perception synthesis is proposed.
Zero-Shot Spatial Layout Conditioning for Text-to-Image Diffusion Models
Guillaume Couairon (Meta AI), Jakob Verbeek (Meta AI)
SegmentationGenerationData SynthesisDiffusion modelImageText
🎯 What it does: A zero-shot method called ZestGuide is proposed, which uses a pre-trained text-to-image diffusion model to accurately control the spatial layout of generated images based on user-provided segmentation masks and free-text descriptions without any additional training.
Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields
Jonathan T. Barron (Google Research), Peter Hedman (Google Research)
GenerationData SynthesisOptimizationComputational EfficiencyNeural Radiance FieldImage
🎯 What it does: This paper proposes Zip-NeRF, which integrates efficient training of Instant Neural Graphics Primitives (iNGP) with the anti-aliasing technology of mip-NeRF 360, achieving a fast and anti-aliasing neural radiance field.
Zolly: Zoom Focal Length Correctly for Perspective-Distorted Human Mesh Reconstruction
Wenjia Wang (SenseTime Research), Taku Komura (Zhejiang University)
Pose EstimationConvolutional Neural NetworkImageMesh
🎯 What it does: A 3D human mesh reconstruction method called Zolly is proposed for perspective-distorted images, which can accurately estimate the camera focal length, the distance between the human body and the camera in close-up images, and reconstruct high-quality 3D human meshes.
zPROBE: Zero Peek Robustness Checks for Federated Learning
Zahra Ghodsi (Purdue University), Farinaz Koushanfar (San Diego State University)
Federated LearningSafty and PrivacyImage
🎯 What it does: A privacy-preserving robust federated learning framework based on zero-knowledge proofs, zPROBE, is proposed, which ensures the privacy of user updates while resisting Byzantine attacks.