Point Cloud Registration as a Classification Problem

Abstract

Point Cloud Registration (PCR) is an important step in fields such as robotic manipulation,
augmented and virtual reality, SLAM, etc. In the context of computer vision, registration
in general refers to the process of aligning data obtained from different frames and as the
name suggests PCR is the task of aligning point-clouds. The main contribution of this
thesis is drawing parallels between PCR and classification which allows us to apply well
studied concepts from classification in PCR. This thesis further shows two applications of
drawing such parallels in the context of deep learning based PCR. We show the use of cross-
entropy loss, and discrepancy loss from classification for partial to full PCR, and outlier
filtering respectively. We finally show that many of the existing deep learning based PCR
architectures can be easily modified to be trained using the loss functions from classification
literature.