Learning-Deep-Learning

Semi-Local 3D Lane Detection and Uncertainty Estimation

March 2020

tl;dr: Patch-based detect-then-cluster 3D lane line detection.

Overall impression

This paper is from the same authors for 3D LaneNet, and improved the performance by a large margin.

The parameterization of the lane lines within each patch is by a polar coordinate (r and $\theta$). The advantage of this approach over directly predicting Cartesian offset from the tile center is unclear. In theory these two representation should be equivalent.

The main advantage of this paper seems to be the introduction of tiles (2D anchors) as compared to 1D anchors in 3D LaneNet. This helps the model generalize to more complex topologies in a simpler way (merges, splits, perpendicular lanes). In contrast, 3D LaneNet deals with merges and splits with regression of two targets per anchor, and cannot handle lane lines vertical tot he ego car.

The paper also introduced uncertainty estimation. How to use this uncertainty and how did this help with the lane detection is not clear.

This paper is republished as 3D LaneNet+ at NeurIPS 2020 workshop.

Key ideas

Technical details

Notes