LAUSR

Semantic bird-eye-view (BEV) grid map is a straightforward data representation for semantic environment perception. It can be conveniently integrated with downstream tasks, such as motion planning, trajectory prediction, etc. Most existing methods of semantic BEV grid-map generation adopt supervised learning, which requires extensive hand-labeled ground truth to achieve acceptable results. However, there exist limited datasets with hand-labeled ground truth for semantic BEV grid map generation, which hinders the research progress in this field. Moreover, manually labeling images is tedious and labor-intensive, and it is difficult to manually produce a semantic BEV map given a front-view image. To provide a solution to this problem, we propose a novel semi-supervised network to generate semantic BEV grid maps. Our network is end-to-end, which takes as input an image from a vehicle-mounted front-view monocular camera, and directly outputs the semantic BEV grid map. We evaluate our network on a public dataset. The experimental results demonstrate the superiority of our network over the state-of-the-arts.