PC-OPT: A SfM Point Cloud Denoising Algorithm

Abstract

AbstractDifferent from 3D models created by digital scanning devices, Structure From Motion (SfM) models are represented by point clouds with much sparser distributions. Noisy points in these representations are often unavoidable in practical applications, specifically when the accurate reconstruction of 3D surfaces is required, or when object registration and classification is performed in deep convolutional neural networks. Outliers and deformed geometric structures caused by computational errors in the SfM algorithms have a significant negative impact on the postprocessing of 3D point clouds in object and scene learning algorithms, indoor localization and automatic vehicle navigation, medical imaging, and many other applications. In this paper, we introduce several new methods to classify the points generated by the SfM process. We present a novel approach, Point-Cloud Optimization (PC-OPT), that integrates density-based filtering and surface smoothing for handling noisy points, and maintains the geometric integrity. Furthermore, an improved moving least squares (MLS) is constructed to smooth out the SfM geometry with varying scales.Keywords3D noise reductionOutlier removingSfMDensity-based clusteringSurface smoothing