Openlane-V2: A Diverse Extension of Openlane-V Based On Low-Rank Representation of Road Lanes.

Authors:

Abstract
We introduce OpenLane-V2, a video lane detection dataset that extends OpenLane-V to include more diverse and challenging scenes, and assess the capability of the eigenlane representation on such complicated scenes. The new dataset includes additional samples with curved lanes, low-light conditions, occlusions, and invisible lanes that can only be inferred from context. We evaluate the recursive video lane detector (RVLD), which is based on the low-rank approximation of lanes (i.e. eigenlanes), on OpenLane-V2 comparatively with other state-of-the-art detectors. The results demonstrate that the eigenlane representation is also effective in handling complex real-world lane detection scenarios.
Funding Information
  • National Research Foundation