Toward Generalized 3D Lane Representation with Lane Geometry Supervision for Autonomous Driving

Lane detection is crucial for autonomous driving. Recent advancements have expanded traditional two-dimensional (2D) lane detection to three-dimensional (3D) by predicting lane positions in 3D space. These methods rely on fully supervised learning, requiring high-quality 3D labels, which are difficult to obtain. This challenge motivates a weakly supervised approach leveraging abundant and easily scalable 2D lane annotations. Specifically, we systematically analyze lane geometric structure priors and introduce Lane Geometry Supervision (LGS), which relies solely on 2D lane labels. Extensive experiments on both synthetic and real-world datasets demonstrate that our method achieves performance comparable to fully supervised approaches using direct 3D labels. Moreover, incorporating LGS as a regularization term further enhances the performance of existing fully supervised methods. Finally, we show that LGS enables a label-efficient training methodology for 3D monocular lane detection, effectively utilizing both scarce yet complete 3D lane labels and abundant but incomplete 2D lane labels.