Automatic detection of road networks from satellite imagery using computer vision

Accurate roadmaps are hard to create and maintain. RoadTracer – a CNN model developed by MIT et al., which constructs road graphs iteratively from satellite images using a convolutional neural network (CNN) as a decision function, was implemented in this project. To evaluate RoadTracer’s implementation, we use two sets of cities: a set that was originally used on the test set in and a set that considers cities with significantly different road network characteristics. Results show that our implementation detected an 8% higher ratio of correct and 31% higher ratio of extra intersections than the original RoadTracer.

Here is the final presentation we made on our implementation, process, and findings:

More information can be found through our technical paper we wrote on the topic. Thanks to the RoadTracer team for baseline which was used in our project.