Unsupervised End-to-end visual odometry by maximizing mutual information

• Grant number: 20/08846-0
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: August 01, 2020
• End date: February 28, 2021
• Field of knowledge: Physical Sciences and Mathematics - Computer Science
• Principal Investigator: Esther Luna Colombini
• Grantee: Rafael Figueiredo Prudencio
• Host Institution: Instituto de Computação (IC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil

Abstract

One of the core objectives of machine learning is to learn useful represen- tations. A simple idea that has not been explored in many fields is to train representation-learning functions to maximize mutual information between its inputs and outputs. This idea is powerful because it enables learning features in an unsupervised manner. Unsupervised learning is extremely important in tasks where it is costly to obtain ground-truth data. In this context, this project aims to apply a mutual information maximization framework to visual odometry, which is the process of estimating the egomotion of an agent using only the input of its camera. Despite being costly to obtain ground-truth per-frame pose information, most attempts to tackle visual odometry still use supervised learning approaches. To the best of our knowledge, mutual information maximization has not been applied to visual odometry as an unsupervised learning approach. Hence, this work proposes a new unsupervised learning framework for visual odometry and aims to quantify its contribution to the task.