CARS: A Photogrammetry Pipeline Using Dask Graphs to Construct A Global 3D Model

CNES and Airbus are currently preparing the CO3D mission [1], aiming at producing global digital surface model (DSM) of the Earth's landmass for both civil and military needs. In this frame, CNES develops CARS, a stereovision pipeline for satellite images, tailored for scalability and robustness. CARS workflow is composed of two main steps: the preparation step which aims at obtaining the best possible epipolar geometry and estimating the disparity range that should be explored, and the DSM computation step, which processes several stereo-rectified pairs in order to produce a single raster DSM. Under the hood, CARS makes use of Dask graph-based and distributed computing of the different steps with lazy evaluation and dynamic task allocation, allowing to run the computation on several tenth of nodes seamlessly. This paper also includes an in-depth example of use.