Weakly-supervised semantic segmentation of rails point cloud data using label diffusion

Abstract

Fugro collects point cloud data, a form of unordered spatial data, with LiDAR, a laser scanning system. Using this type of data, detailed maps of the surrounding of railway tracks can be generated. In order to do this, it is necessary to label all the points in these collected point clouds. This can be done in multiple ways. In this thesis, we explore the usefulness of Label Diffusion LiDAR Segmentation (LDLS). We use a panoptic segmentation model on Fugro video data to perform semantic segmentation and verify its performance on RailSem19 data, an open source semantic segmentation dataset of railway image data. LDLS uses the output of this model to diffuse the labels through the point cloud. We show multiple visual results of a few different configurations. We also give quantitative results of the performance for a subset of platform points.