Comparison of point cloud densification from multi-view stereo and 3D Gaussian splatting in industrial photogrammetry

Photogrammetry is increasingly popular in the modern manufacturing industry, for performing non-contact fast optical measurements of manufactured components. During photogrammetry, the goal is to obtain accurate three-dimensional reconstructions of a measurand that include information about surface form and dimensions. Using multiple two-dimensional observations of a measurand, a 3D reconstruction of the measurand can be constructed algorithmically. The main objective of this research is to compare the outcomes of the 3D reconstruction dense point cloud obtained by 3D Gaussian splatting (3DGS) to the general multi-view stereo (MVS) method. Using the same set of 2D images of the measurand as the input to the measurement pipelines of the two aforementioned methods, we obtained the full position and orientation (i.e. the extrinsic parameters) of the camera in a world coordinate system by calibrating the camera's pose for each image taken. The proposed study involves the introduction of a camera extrinsic parameter estimation step for each of the methods to obtain a point cloud that is reconstructed in the actual physical size of the measurand. The geometric dimensions and surface form details of the measurand were also measured with fringe projection profilometry (FPP) techniques using a commercial instrument GOM ATOS, presented here as the ground truth measurement to compare the scale-calibrated point clouds obtained by each of the methods above for the accuracy evaluation of the measurements.