1. Hierarchical Recurrent Attention Networks for Structured Online Maps

• https://patrick-llgc.github.io/Learning-Deep-Learning/paper_notes/hran.html

预测结果：

• 两个点数不对称的折线

Loss

• Polyline Loss：首先有两个点集P和Q，然后分别取P、Q每个点距离另一个polyline所有pixel中最近pixel的距离作为loss。

2. DAGMapper: Learning to Map by Discovering Lane Topology

• https://github.com/patrick-llgc/Learning-Deep-Learning/blob/master/paper_notes/dagmapper.md

预测结果：

• 两个点数不对称的折线

Loss

• symmetric Chamfer distance（对称倒角距离）：首先密集采样两个点集P和Q，然后分别取P、Q每个点距离另一个点集中的最短距离作为loss。
• cosine similarity loss（余弦相似性）：通过该loss来学习方向；
• L2 loss：学习距离变化。

参考文章：

• https://github.com/patrick-llgc/Learning-Deep-Learning/blob/master/paper_notes/polymapper.md
• https://patrick-llgc.github.io/Learning-Deep-Learning/paper_notes/dagmapper.html