From October 2020 to April 2021, I completed an internship in the Computer Vision group of Professor Daniel Cremers at the Technical University of Munich. The goal of the project was to use E(2)-equivariant CNNs to predict cloud types from CUMULO, a dataset of hyperspectral images with sparse cloud labels.
During this project, I extended my personal deep learning repertoire with equivariant architectures and learned how to train models on geospatial datasets. Pleasant side results of this project were artfully colored satellite images like the one shown above. The E(2)-equivariant CNNs reached a classification accuracy near that of the state-of-the-art, but the semester and my internship ended before any meaningful results could be published.