Road Surface Segmentation .

Abstract:

Road surface segmentation is a crucial aspect of self-driving cars, autonomous robots, and advanced driving assistant systems, which attracted much research attention to this topic. Deep learning models are being developed with the availability of large road segmentation datasets. However, the diversity of data obtainable is finite and it is not transparent to what degree techniques generalize beyond the explicit datasets they were trained on. In this paper, we propose a novel training process to improve the cross-dataset performance of road segmentation systems by forcing the model to learn features that are highly relevant to the output required.