Evaluation of synthetically generated patterns for image-based 3D reconstruction of texture-less objects

Abstract

© 2019 Elsevier Ltd Every material and every surface has its own visual texture. Surfaces with monotone, repetitive or uniform visual texture represent a challenge for the image-based Structure From Motion Multi-View Stereo three-dimensional (3D) reconstruction method. It is possible to overcome the lack of visual texture on the surfaces by projecting synthetically generated images (patterns) using a video projector. This research proposes the generation of the synthetic images that are based on digits of irrational numbers pi, phi, e, sqrt2, sqrt3 and digits produced by the random number generator. Images were divided into three classes based on the number of hues. The aim is to evaluate synthetically generated images and determine the characteristics of the most suitable synthetic image(s) that increase the accuracy of the final reconstructed polygonal 3D model. The synthetically generated images were evaluated using results of the multi-criteria analysis as well as real and virtual planar surface 3D digitization methods, where images with uniform distribution histogram have the most suitable characteristics. To verify evaluation results, 3D reconstruction of aluminium test model was carried out. Four polygonal 3D models of the aluminium test model were reconstructed. Three polygonal 3D models were reconstructed using projected patterns of each class, and one was reconstructed in daylight condition. The resulting accuracies of reconstructed polygonal 3D models were evaluated using the Computer-Aided Inspection. The polygonal 3D model obtained by projecting class I pi I pattern with the strong random stochastic visual texture achieved std. distance of 0.173 mm and mean distance of 0.016 mm compared to the polygonal 3D model obtained in daylight condition with std. and mean distance of 1.188 mm and −0.139 mm respectively.