One application is related to the update procedure of an existing map database. The most time-consuming task in this process is to detect changes between the update and the existing database. New buildings may have been built and new roads could have been created. The goal of this work is to detect changes automatically by using deep learning.

However, change detection in such complex image data is not a trivial task and require large amounts of training data that is properly labeled. Our project with change detection for map updates is expected to have a reasonable amount of image data but a limited amount of labelled data. We will investigate and develop models with priors, like traditional feature extraction, and combine them with neural networks, and semi-supervised methods.

A part of the work will also be to investigate the need for further allocation of labelled data. The scarcity in labelled data might be most critical for automatic change detection, and it is likely that this activity will run over several years.