A comparative study of semantic segmentation of omnidirectional images from a motorcycle perspective

The semantic segmentation of omnidirectional urban driving images is a research topic that has increasingly attracted the attention of researchers, because the use of such images in driving scenes is highly relevant. However, the case of motorized two-wheelers has not been treated yet. Since the dynamics of these vehicles are very different from those of cars, we focus our study on images acquired using a motorcycle. This paper provides a thorough comparative study to show how different deep learning approaches handle omnidirectional images with different representations, including perspective, equirectangular, spherical, and fisheye, and presents the best solution to segment road scene omnidirectional images. We use in this study real perspective images, and synthetic perspective, fisheye and equirectangular images, simulated fisheye images, as well as a test set of real fisheye images. By analyzing both qualitative and quantitative results, the conclusions of this study are multiple, as it helps understand how the networks learn to deal with omnidirectional distortions. Our main findings are that models with planar convolutions give better results than the ones with spherical convolutions, and that models trained on omnidirectional representations transfer better to standard perspective images than vice versa.