CARS

Two eyes judge depth from how far a thing shifts between them, and two satellite images taken from different angles do the same for the ground.³ From such pairs CARS builds a digital surface model, a raster grid in which each cell holds the height of the surface seen from above.^2,3 A grid of heights is not a full 3D mesh, so it is often called 2.5D.³

CNES maintains it on GitHub.¹ It is the core of the 3D image processing chain for the CO3D satellite mission, and the AI4GEO and Evoland projects use it too.^1,2 It handles large production runs on a computing cluster and also runs on a personal computer.² The current pipeline is described in a 2026 paper in the ISPRS Archives by D. Youssefi and colleagues.¹ Release 1.3.0 came out in October 2026.⁶

The work goes in four steps. First the images are resampled so that a point can move only along one line between them, and the search for a match becomes one-dimensional.³ Matching then finds each pixel's partner in the other image.³ Crossing the two lines of sight places the point in space, with a longitude, latitude and altitude.³ Last, every point is projected onto a grid of heights and colours.³ Any image GDAL can read will do, as long as it comes with RPC coefficients, a compact formula linking image pixels to ground positions.³ Example setups exist for Pléiades, SPOT 6 and 7, Pléiades Neo and Maxar WorldView.⁵

Matching can fail on moving objects such as vehicles, in shadows, where one viewpoint is blocked, and on surfaces with too little texture.³ Installing takes pip install cars on Linux or the cnes/cars Docker image.^1,4 The quick start runs on a free Pléiades sample of the pyramids at Giza.¹

For a model of a city or a valley, CARS gives the height of the ground and what stands on it from the stereo images the user brings. The code is under the Apache 2.0 licence.¹