Hierarchical Recurrent Attention Networks for Structured Online Maps

TL;DR

This paper introduces a hierarchical recurrent attention network for extracting structured road maps from sparse 3D point clouds, mimicking human annotation by sequentially identifying and drawing lane boundaries.

Contribution

It proposes a novel hierarchical recurrent network with a differentiable loss function tailored for structured polyline prediction in road network extraction.

Findings

Achieves 92% accuracy in recovering correct topology on highway data.

Effectively traces lane boundaries by attending to initial regions and completing polylines.

Demonstrates improved structured map extraction from sparse 3D data.

Abstract

In this paper, we tackle the problem of online road network extraction from sparse 3D point clouds. Our method is inspired by how an annotator builds a lane graph, by first identifying how many lanes there are and then drawing each one in turn. We develop a hierarchical recurrent network that attends to initial regions of a lane boundary and traces them out completely by outputting a structured polyline. We also propose a novel differentiable loss function that measures the deviation of the edges of the ground truth polylines and their predictions. This is more suitable than distances on vertices, as there exists many ways to draw equivalent polylines. We demonstrate the effectiveness of our method on a 90 km stretch of highway, and show that we can recover the right topology 92\% of the time.