Analysing the quality of PlaneRCNN plane parameters using a structure from motion system

Abstract: This report aims to compare the plane parameters given by the PlaneRCNN network with those created by using a Structure from Motion (SfM) solution on the same data. Several datasets are used, and the results from PlaneRCNN are evaluated by using the camera matrices from the SfM solution and applying these along with scaling and coordinate transformations. We use and explain RANSAC, neural networks of different types, and general computer vision background material (such as methods of plane creation) required to understand the component parts of the report. A general analysis of the PlaneRCNN network is offered, using different performance metrics on varying sets of data, along with practical advice on how to best use it. The end result is that we find a good method for comparing the PlaneRCNN plane parameters to those generated by SfM systems. We also note under what conditions we found PlaneRCNN to give the best results.