2023-03-07, 11:45–12:15, 01.S.21 (1st floor)
A run through a PyTorch-based tool for generating clouds in satellite images.
This talk will outline the key challenges associated with the presence of clouds in satellite imagery in the context of training and evaluating machine learning models.
As one part of the solution, the tool for efficient simulation of clouds and cloud shadows in satellite imagery is presented here. Specific features of the SatelliteCloudGenerator will be outlined and demonstrated, with tips on how to control the appearance of the generated clouds.
Finally examples of how simulated clouds can be used for both training and evaluating satellite image processing models will be shown.
⚙️ All code presented during the talk will be available publicly as Google Colab notebooks. Software used for the demonstration will be based in PyTorch, PyTorch Lightning, Numpy.