Automated Data Generation for Training of Neural Networks by Recombining Previously Labeled Images

Abstract

In this paper, we present our approach to data generation for the training of neural networks in order to achieve semantic segmentation in an autonomous environment. Using a small set of previously labeled images, this approach allows to automatically increase the amount of training data available. This is achieved by recombining parts of the images, while keeping the overall structure of the scene intact. Doing so allows for early network training, even with only few training samples at hand. Furthermore, first results show that training networks using the so created datasets allow for good segmentation results when compared to publicly available datasets.