PersFormer_3DLane

Why we included this project

Most lane detectors only work in the 2D image plane, which does not give a self-driving stack the geometry it needs to actually plan around lane boundaries. PersFormer instead takes a single front camera frame and lifts it into a bird's-eye view using a Transformer module that matches front-view regions to ground-plane locations, then predicts lanes in 3D space. Its unified 2D/3D anchor design and auxiliary task mean the model produces both image and BEV lane outputs at once, keeping the two consistent without separate systems. The repo includes training and evaluation code, pretrained weights, and benchmark numbers on OpenLane and ONCE_3DLanes, so you can reproduce the reported results and then adapt the model to your own camera setup or dataset. For a team starting on lane perception, that makes it a practical baseline to extend rather than a paper to reimplement from scratch.