Instance Segmentation Applied to Underground Infrastructures

As underground infrastructures suffer from strikes during the ground excavation procedure resulting from poor subsurface utilities documentation, this paper aims to improve the mapping of such elements by integrating instance segmentation techniques. To perform supervised training of the well-known Mask R-CNN architecture, we created our own dataset based on surveys that we have access to from the SYSLOR company, resulting in around 2600 labelled images. Through several training sessions, performed with K-fold cross validation, we studied the level of contribution of each construction sites selected for the dataset creation. Currently we achieved a very good 38.4 mean average precision (mAP) on three defined classes: sheaths, twisted pipes and smooth pipes.