August 2020
tl;dr: Dynamic training of a CNN as an DRL agent to draw maps.
The following work are focused on road network discovery and are NOT focused on HD maps.
- DeepRoadMapper: semantic segmentation
- RoadTracer: like an DRL agent
- PolyMapper: iterate every vertices of a closed polygon
RoadTracer noted the semantic segmentation results are not a reliable foundation to extract road networks. Instead, it uses an iterative graph construction to get the topology of the road directly, avoiding unreliable intermediate representations.
The network needs to make a decision to step a certain distance toward a certain direction, resembling an agent in a reinforcement learning setting. This is somehow similar to the cSnake idea in Deep Boundary Extractor.
- Summary of technical details
- Questions and notes on how to improve/revise the current work