Labeling Custom Indoor Point Clouds Through 2D Semantic Image Segmentation

Abstract

For effective Computer Vision (CV) applications, one of the difficult challenges service robots have to face concerns with complete scene understanding. Therefore, various strategies are employed for point-level segregation of the 3D scene, such as semantic segmentation. Currently Deep Learning (DL) based algorithms are popular in this domain. However, they require precisely labeled ground truth data. Generating this data is a lengthy and expensive procedure, resulting in a limited variety of available data. On the contrary, the 2D image domain offers labeled data in abundance. Therefore, this study explores how we can achieve accurate labels for the 3D domain by utilizing semantic segmentation on 2D images and projecting the estimated labels to the 3D space via the depth channel. The labeled data may then be used for vision related tasks such as robot navigation or localization.