About the Ambiguity of Data Augmentation for 3D Object Detection in Autonomous Driving
Matthias Reuse, Martín Simón, Bernhard Sick
Abstract
Although data augmentation is considered an important step in the training strategy of 3D object detectors on point clouds to increase the overall performance and robustness, in almost all publications the topic of augmentation and the choice of the individual augmentation methods used are only addressed very briefly with reference to previous work and are not backed up with sufficient experiments. The question therefore arises as to the impact and the transferability of different augmentation policies. Through a series of elaborate experiments with four networks on two datasets, this paper shows that the positive effects of different data augmentation methods are not so clear-cut and instead depend strongly on the network architecture and the dataset.