Robust Point Cloud Reconstruction Based On Self-supervised Learning

3D reconstruction is a key problem of computer graphics and computer vision,with growing demand for 3D digital models in various industries such as gaming,robotics,and art creation.Conventional methods for creating digital models are time-consuming and require trained professionals,presenting a challenge for usual users.There are two main types of 3D models:object-level and scene-level,and a high-quality reconstruction result requires accurate surface representation as well as properties such as smoothness and simplicity.Despite advances in scanning technology,point clouds produced by the hardware may contain defects such as unoriented,noise,missing data,uneven distribution,and density variety,leading to inaccuracies and instability in the reconstruction results.To overcome these challenges,this paper presents a robust point cloud reconstruction algorithm based on self-supervision learning.By considering the implicit moving leastsquares algorithm and the network as a dual representation of the surface,the algorithm trains the neural network through mutual regularization.The neural network provides point normals for the implicit moving least-squares algorithm(IMLS).At the same time,we can get the pseudo-label from the IMLS for training the neural network,helping to recover shape features near the surface.The results of numerous experiments show that this self-supervised mutual regularization strategy can accurately reconstruct surfaces and effectively handle different levels of noise and other defects without supervision and normals.The proposed algorithm demonstrates its potential in practical applications through experiments on challenging real scans.