Encoding A Mathematically Faithful DeepVIO Solution

Abstract

Within the last decade visual odometry (VO) has been continually accumulating the research interests of the computer vision community. The accent of artificial intelligence (AI) is in real-time rerouting the interest of researchers from the traditional feature point methods, to AI based solutions - primarily those based on deep learning (DL), which in turn has forced the VO literature to become increasingly opaque. In an attempt to strike a balance between the understandability/robustness of new model and the ever-increasing temptation of exceedingly sophisticated black box models, this paper produced a highly sophisticated DL based End-to-End VO solution which religiously encodes classical mathematical VO solutions. This paper has developed a method to encode a mathematically proven traditional VO solution into a DL based solution with high transparency and achieved results inline with those of the comparable solution on the KITTI leaderboard.