Road Semantic Segmentation Oriented Dataset for Autonomous Driving

Abstract

There are various datasets for segmentation such as KITTI, CamVid, Cityscapes, ApolloScape, Mapillary Vistas, etc. These datasets try to include as many classes as possible, and classify objects in pixels across all regions of an image. However, the greater the number of classes, the more difficult it is to learn through deep learning. In particular, from an autonomous driving point of view, it is more important to understand the road situation than to detect the sky or buildings. This paper proposes a Road Semantic Segmentation Oriented (RSSO) Dataset. This dataset divides the various markers into eight classes, including lanes on the road. ApolloScape's lane segmentation dataset is also provided by labeling only road markers. However, RSSO dataset has the advantage of finding drivable regions by including asphalt in the class. We expect the proposed dataset to be useful for a variety of autonomous driving applications.