A Visual Method to study the Error Function of ICP Algorithms

Abstract

This paper proposes a novel method to study the error function of ICP based algorithms. With our method, we visualize the multidimensional error function of ICP algorithms in one dimension which allows us to compare and quantify the performance of ICP algorithms in an intuitive and descriptive manner. This is motivated by the fact that there are many ICP variants around and for researchers and engineers it is challenging which algorithm they shall choose. New approaches are often only evaluated based on runtime and accuracy. Our visual method allows to gain further insights beyond those metrics. We demonstrate the capability of our method by applying it to the KITTI LIDAR odometry benchmark. Our experiments show evidence for errors in the ground truth data, difficulties in highway scenarios and prove the power of superior error metrics such as the newly emerged symmetric objective function.