Generating Training Data for Deep Learning-Based Segmentation Algorithms by Projecting Existing Labels onto Additional Aerial Images

Kurzfassung

Highly accurate manually-generated labels in aerial and satellite images are used for the training of deep learning-based segmentation algorithms and should be available in large numbers and cover many different scenarios to increase the accuracy and generalization capability of the underlying models. Existing labels can be efficiently reused by photogrammetric projections onto additional overlapping aerial or satellite images, enabling great variability in the appearance of the scenes based on differences in viewing angles and environmental conditions. In this work, we investigate whether the additionally generated training data can effectively lead to an increase in prediction accuracy. To this end, we collected aerial images overlapping with the already annotated Traffic Infrastructure and Surroundings (TIAS) dataset, taken from a large-scale historical database spanning 2011 to 2024, and generated new training data by means of photogrammetric projections of existing labels onto these additional images. Training a Dense-U-Net model on the whole TIAS dataset or a part therefore, with and without additional projected labels, showed that this technique could be beneficial to improve the performance of a model if only a small amount of annotations is available comparatively to a large amount of overlapping aerial images.