Generation of Datasets for Semantic Segmentation from 3D Scanned Data to Train a Classifier for Visual Navigation

Abstract

The authors attempt to actualize robot navigation using only visual information and propose a novel scheme based on the results of semantic segmentation for road-following in urban and indoor scenes. It is shown that the proposed vision-based navigation performs well when accurate results of semantic segmentation are obtained. Generally, to construct a good classier, an appropriate dataset should be prepared. However, creating a dataset for semantic segmentation with pixel-wise class labels to all pixels involves a significant amount of human effort. Hence, we proposes a novel approach to generate datasets for semantic segmentation that is suitable for the visual navigation of a robot. The proposed approach provides two-dimensional images with pixel-wise class labels from three-dimensional scanned data. Experimental results show that the generated dataset can be used for training semantic segmentation using ICNet.