Anchor3DLane: Learning to Regress 3D Anchors for Monocular 3D Lane Detection

Monocular 3D lane detection is a challenging task due to its lack of depthinformation. A popular solution is to first transform the front-viewed (FV)images or features into the bird-eye-view (BEV) space with inverse perspectivemapping (IPM) and detect lanes from BEV features. However, the reliance of IPMon flat ground assumption and loss of context information make it inaccurate torestore 3D information from BEV representations. An attempt has been made toget rid of BEV and predict 3D lanes from FV representations directly, while itstill underperforms other BEV-based methods given its lack of structuredrepresentation for 3D lanes. In this paper, we define 3D lane anchors in the 3Dspace and propose a BEV-free method named Anchor3DLane to predict 3D lanesdirectly from FV representations. 3D lane anchors are projected to the FVfeatures to extract their features which contain both good structural andcontext information to make accurate predictions. In addition, we also developa global optimization method that makes use of the equal-width property betweenlanes to reduce the lateral error of predictions. Extensive experiments onthree popular 3D lane detection benchmarks show that our Anchor3DLaneoutperforms previous BEV-based methods and achieves state-of-the-artperformances. The code is available at:https://github.com/tusen-ai/Anchor3DLane.