Learning-Deep-Learning

LaneNet: Towards End-to-End Lane Detection: an Instance Segmentation Approach

March 2020

tl;dr: Binary segmentation and learning of embedding for instance segmentation, plus learned perspective mapping (adaptive IPM).

Overall impression

The paper proposed a good method to perform instance segmentation on long/thin objects as conventional detect and segment pipeline does not work well (they are better suited for compact objects). Binary segmentation + embedding for clustering into instances.

The clustering idea directly inspired Semilocal 3D Lanenet to cluster image tiles together which the same lanes pass through.

The idea to predict vanishing point to guide laneline detection is similar to VPGNet, but LaneNet is not predicting a point but rather directly predicting the homographic transformation.

The HNet is trained separately. This decoupled design makes the system scalable. Similar to the idea of the 3DGeoNet in Gen-LaneNet.

Key ideas

Technical details

Notes