Dishpolish: Exploring the Recovery of Geometric Invariants in Deep Learning Models via Pose Estimation of Microwave Dish Antennae

In this paper, we propose a deep learning algorithm, called Dishpolish to exploit geometric invariances of microwave dish antennae to perform a dish mensuration task-the study of estimating a microwave antenna dish pointing direction from imagery. We synthesize a dish imagery dataset from 5 distinct dish models and evaluate it on the proposed Dishpolish deep-learning architecture. We find that the method is capable of recovering dish orientation to 5.36° and 1.03° azimuth and elevation error, respectively, when it has seen the dish model before, but performs less effectively on unseen dishes.