Dense Point Cloud Quality Factor as Proxy for Accuracy Assessment of Image-Based 3D Reconstruction

Abstract

Photogrammetry using structure from motion (SfM) and multiview stereopsis (MVS) techniques can recover three-dimensional (3D) structure from a set of overlapping, unoriented, and uncalibrated images captured by nonmetric digital cameras. It is possible to generate accurate reconstructions of sparse points using mathematically robust bundle adjustment procedures together with accurate surveying control data. However, MVS, which recovers the dense geometry by matching and expanding between sparse points, is prone to additional error. Miscellaneous constituents such as sensor specifications, data collection, and site conditions can introduce random noise or artifacts that locally degrade the accuracy of the dense point cloud. This paper proposes seven indexes, named dense point cloud quality factors (DPQFs), as proxy indicators of image-based dense reconstruction accuracy. DPQFs include proximity to keypoint features, distance to GCPs, angle of incidence, camera stand-off distances, number of overlapping images, brightness index, and darkness index. The correlation between the DPQFs and the 3D error was investigated in simulated and empirical experiments scenarios with varying factors. The results of this study showed that the DPQFs provide proxy indications for accuracy when the error estimation for the dense point clouds is more challenging than error propagation computations in bundle adjustment (BA). The DPQFs can be defined solely using the SfM-MVS data, without prior knowledge about the error. Inclusion of the factors as additional fields of information and their visualization provide tangible intuitions regarding the factors that influence the accuracy of image-based 3D reconstruction.