Master’s thesis, in a collaboration with Aarhus University, Denmark
Deep learning has led to the development of end-to-end methods for visual odometry (VO), setting up a new paradigm for this family of algorithms. The higher-level representation of the environment that deep learning introduces allows it to generate estimates under challenging imaging conditions that make traditional algorithms fail. However, their performance is conditioned by the selected training environment. This limitation has been mainly addressed by feeding more and more varied data to the network. The present work aims to assess the information in the datasets from which the algorithm is learning. For that, a set of metrics to deploy on the target datasets will be developed. The aim of these metrics is twofold: firstly, to assess the quality and variability of the data available, and secondly, to establish a criterion for the data splitting in learning-based VO.
Contact: Erdal Kayacan