RoadTracer Code

This is the code for "RoadTracer: Automatic Extraction of Road Networks from Aerial Images".

There are several components, and each folder has a README with more usage details:

• dataset: code for dataset preparation
• roadtracer: RoadTracer
• roadcnn: our segmentation approach (baseline)
• deeproadmapper: DeepRoadMapper (baseline)

You will need gomapinfer (https://github.com/mitroadmaps/gomapinfer/) as a dependency.

The training/inference code is built on top of TensorFlow.

Usage

First, follow instructions in dataset/ to download the dataset.

Then, follow instructions in the other folders to train a model and run inference.

Junction Metric

The junction metric matches junctions (any vertex with three or more incident edges) between a ground truth road network graph and an inferred one.

go run junction_metric.go /data/graphs/chicago.graph chicago.out.graph chicago

Visualization

viz.go will generate an SVG from a road network graph. It will refer to the /data/testsat/ images; to view the SVG, those images will need to be in the same folder as the generated SVG.

go run viz.go chicago /data/graphs/chicago.graph
go run viz.go chicago chicago.out.graph

Applying RoadTracer on a new region

You need to make a few modifications to run the code on a region outside of the 40-city RoadTracer dataset.

First, download the imagery. Update dataset/lib/regions.go and put a latitude/longitude bounding box around your region in the regionMap. You can comment out the existing regions. Then, follow instructions in dataset/ for running 1_sat.go.

Then, update roadtracer/tileloader.py and set TRAINING_REGIONS and REGIONS to just a list with your region label from the regionMap. Also update REGION, TILE_START, and TILE_END in infer.py (e.g. if your imagery tiles were saved as xyz_-1_-1_sat.png through xyz_1_1_sat.png, set TILE_START = -1, -1 and TILE_END = 2, 2.

Finally, manually specify a starting location for RoadTracer in infer.py. Set USE_TL_LOCATIONS = False, MANUAL_RELATIVE = 0, 0, and set the manual points to the pixel positions of two points on the road network in xyz_0_0_sat.png. These two points should be close to each other (around SEGMENT_LENGTH apart) and best to be in the middle of a road.

Now you should be able to get a road network graph by running infer.py.

To convert this to latitude/longitude, you can use dataset/convertarg.go:

go run convertarg.go YOUR_REGION_LABEL frompix out.graph out.lonlat.graph