Map Inference from Satellite Segmentation Data through Reinforcement Learning: A Novel Approach

Abstract

Online road maps need to be kept up to date for a variety of purposes, and the task of updating them can be automated. There are many algorithms to infer road map structure from data, including satellite imagery and crowdsourced GPS trajectories. However, most of these algorithms use supervised learning and require hyperparameter tuning on a given location to be able to infer maps with high accuracy. In addition, these algorithms are trained for metrics like per-pixel loss but not trained on end-to-end objectives. In this project, we experiment with a Reinforcement Learning based algorithm that may counter the limitations of current algorithms.

We use a map extraction algorithm with heuristics as a baseline and demonstrate that our RL algorithm achieves precision and recall that are comparable to the baseline algorithm. The RL algorithm is able to do this without much hyperparameter tuning, whereas the baseline requires aggressive hyperparameter tuning to give comparable results. In addition, the RL agent can be trained end-to-end to directly maximize the relevant metrics, including the topology of the extracted road network, whereas the baseline requires heuristic post processing to produce such outputs.