面向无监督学习的三维点云补全综述OA

3D point cloud completion represents a significant research avenue within the field of computer vision.Tradi-tional supervised methods for point cloud completion depend heavily on paired data,which limits their generalization capa-bilities.In contrast,unsupervised point cloud completion methods offer the advantage of eliminating the need for manual annotation,thereby enhancing the universality and adaptability of point cloud completion techniques.This paper system-atically reviews the advancements in unsupervised 3D point cloud completion from 2018 to 2025 and categorizes existing methodologies into four distinct types:methods based on unpaired learning,which achieve completion through latent spa-tial alignment or knowledge transfer from unpaired data;methods based on generative models,which employ GAN and diffusion models to generate point clouds characterized by reasonable structure and rich detail;methods based on self-training learning,which leverage the internal geometric rules or pseudo-label signals of point clouds for self-training opti-mization;and cross-modal methods,which integrate multimodal information such as images,depth maps,and text to enhance completion accuracy and robustness.Additionally,this paper introduces prominent datasets such as ShapeNet,KITTI,and ScanNet,outlines evaluation metrics including CD,EMD,F-Score,and MMD,and compares the performance of various algorithms on both synthetic and real datasets.Finally,the paper summarizes the challenges encountered in unsupervised point cloud completion and discusses future research trends.