The training set contains 7,246 labelled aerial image tiles from Freetown, Sierra Leone; Nairobi, Kenya; and Stellenbosch, South Africa. Train.csv contains the OpenStreetMap road centrelines and pixel scale for each tile. Some training tiles contain no mapped roads. The test set contains 4,548 tiles.
External datasets
SpaceNet 3 Roads and SROADEX are external sources you can use in line with the challenge rules. If you use external data or pretrained models, disclose their sources.
Test imagery
Test images are supplied as 256 × 256 pixel, 8-bit RGB PNGs without georeferencing. The ID in Test.csv matches the image filename without the .png extension. The metres_per_pixel column gives the ground distance represented by one pixel. You can use this value to bring images to a common ground scale during preprocessing. Submit your predictions in the original tile’s pixel coordinates.
Files
Extract both archives into the same directory to use the starter notebook with its default data path.
Predictions must be lines in tile pixel coordinates (MULTILINESTRING geometry).
Join the largest network for
data scientists and AI builders