Deep Visual Teach and Repeat on Path Networks

Tristan Swedish, Ramesh Raskar

Abstract: We propose an approach for solving Visual Teach and Repeat tasks for routes that consist of discrete directions along path networks using deep learning. Visual paths are specified by a single monocular image sequence and our approach does not query frames or image features during inference, but instead is composed of classifiers trained on each path. Our method is efficient for both storing or following paths and enables sharing of visual path specifications between parties without sharing visual data explicitly. We evaluate our approach in a simulated environment, and present qualitative results on real data captured with a smartphone.