Abstract

Methods for 3D lane detection have been recently proposed to address theissue of inaccurate lane layouts in many autonomous driving scenarios(uphill/downhill, bump, etc.). Previous work struggled in complex cases due totheir simple designs of the spatial transformation between front view andbird's eye view (BEV) and the lack of a realistic dataset. Towards theseissues, we present PersFormer: an end-to-end monocular 3D lane detector with anovel Transformer-based spatial feature transformation module. Our modelgenerates BEV features by attending to related front-view local regions withcamera parameters as a reference. PersFormer adopts a unified 2D/3D anchordesign and an auxiliary task to detect 2D/3D lanes simultaneously, enhancingthe feature consistency and sharing the benefits of multi-task learning.Moreover, we release one of the first large-scale real-world 3D lane datasets:OpenLane, with high-quality annotation and scenario diversity. OpenLanecontains 200,000 frames, over 880,000 instance-level lanes, 14 lane categories,along with scene tags and the closed-in-path object annotations to encouragethe development of lane detection and more industrial-related autonomousdriving methods. We show that PersFormer significantly outperforms competitivebaselines in the 3D lane detection task on our new OpenLane dataset as well asApollo 3D Lane Synthetic dataset, and is also on par with state-of-the-artalgorithms in the 2D task on OpenLane. The project page is available athttps://github.com/OpenPerceptionX/PersFormer_3DLane and OpenLane dataset isprovided at https://github.com/OpenPerceptionX/OpenLane.