Influence of domain shift factors on deep segmentation of the drivable path of an autonomous vehicle

One of the biggest challenges for an autonomous vehicle (and hence the WEpod) is to see the world as humans would see it. This understanding is the base for a successful and reliable future of autonomous vehicles. Real-world data and semantic segmentation generally are used to achieve full understanding of its surroundings. However, deploying a pretrained segmentation network to a new, previously unseen domain will not attain similar performance as it would on the domain where it is trained on due to the differences between the domains. Although research is done concerning the mitigation of this domain shift, the factors that cause these differences are not yet fully explored. We filled this gap with the investigation of several factors. A base network was created by a two-step fine-tuning procedure on a convolutional neural network (SegNet) which is pretrained on CityScapes (a dataset for semantic segmentation). The first tuning step is based on RobotCar (road scenery dataset recorded in Oxford, UK) while afterwards this network is fine-tuned for a second time but now on the KITTI (road scenery dataset recorded in Germany) dataset. With this base, experiments are used to obtain the importance of factors such as horizon line, colour and training order for a successful domain adaptation. In this case the domain adaptation is from the KITTI and RobotCar domain to the WEpod domain. For evaluation, groundtruth labels are created in a weakly-supervised setting. Negative influence was obtained for training on greyscale images instead of RGB images. This resulted in drops of IoU values up to 23.9% for WEpod test images. The training order is a main contributor for domain adaptation with an increase in IoU of 4.7%. This shows that the target domain (WEpod) is more closely related to RobotCar than to KITTI.