Automatic Detection of Windows Reflection or Transparency Pollution in TLS Acquisitions

Abstract

Three-dimensional acquisitions have been more and more used in recent years, for multiple applications, such as cultural heritage preservation. When these point clouds are generated through laser scanning, transparent and/or reflective objects such as windows can generate inexact or undesirable data. These must be cleaned up by a human, which is often time-consuming and requires experience. This work provides an insight on some methods that can be used to automate this task. It investigates the usage of Mask R-CNN with intensity images in equirectangular projections. The huge images are tiled into squares of 2048x2048 pixels for both training and prediction. The model has good performances on the test and validation sets to handle both types of problems; but also to manage the presence of a mirror in a scene.