Relative Pose Regression using Non-Square CNN Kernels: Estimation of translation, rotation and scaling between image pairs with custom layers

Abstract

Localisation is a research field where handcrafted and complex engineering methods have so far given the best results. However, a general trend in computer science is that data-driven approaches often outperform classical methods. These datadriven approaches have been enabled by better high-resolution range sensors and vast amounts of data. End-to-end deep neural network approaches have much better scalability, but the state of the art localisation networks still use designs originally developed for image classification. These are necessarily not the best-suited design, as they aim to achieve a different task than relative pose regression. We show that custom input layers designed to predict translation, rotation or scaling between images are a possible solution. The networks tested in this thesis give good results, but do not outperform a baseline neural network with the same design as a neural network initially made for image classification.