LaneDetection_End2End

Why we included this project

This is a compact reference implementation of a specific lane-detection idea from an ICCV 2019 workshop paper: instead of the usual two-step segmentation-then-fit pipeline, a single end-to-end network uses a differentiable least-squares layer to regress lane-line curve parameters directly. The repo includes two comparable variants, one of which adds a backprojection loss that regresses toward coordinates in the original perspective view and runs multi-lane experiments on the full TuSimple dataset. That makes it handy as a baseline for benchmarking and as a worked example of making a geometric fitting step differentiable for backpropagation. One thing to keep in mind: the code is under a non-commercial Creative Commons license.