The paper Tractable Large-Scale Deep Reinforcement Learning by Nima Sarang and Charalambos Poullis has been accepted for publication in the Computer Vision and Image Understanding journal. The code is publicly available at https://github.com/nsarang/road-extraction-rl. Some results are shown below.
TL;DR: We frame the problem of road extraction as an RL problem where an agent is trained to traverse the roads from satellite images. We introduce novel techniques for making this inherently complex problem tractable on a minimal number of GPUs and propose a self-supervised loss and actions tailored for road extraction.
The orange lines show the ground truth and are merely there for visualization purposes. The first three videos are from the test dataset and the final from the training.