Learning-Deep-Learning

LSTR: End-to-end Lane Shape Prediction with Transformers

November 2020

tl;dr: Use transformer to directly predict polynomial coefficients of lanes.

Overall impression

The paper uses transformers for lane line detections. It is inspired by DETR, which introduced the transformers to object detection.

The paper has a great session talking about the polynomial lane shape model with very detailed derivation.

The formulation of lane lines as parallel polynomials is a bit limiting as it cannot handle more complex topologies such as splits, merges and lanes perpendicular to the ego lane. However the idea is still applicable if we allow a more flexible representations of lane lines, as long as there is still the concept of individual countable lane line instances (number of query sequence).

The paper recycles a lot of the details from DETR but describes them differently. It is recommended that these two papers should be read together.

Key ideas

Technical details

Notes