Cube it: Training Hyperspectral Demosaicing Models Using Synthetic Datasets

Hyperspectral demosaicing aims to recover full spectral information at each pixel from a mosaicked image captured using a snapshot camera. Cameras vary in terms of used multispectral filter arrays (MSFA). SOTA demosaicing al-gorithms are evaluated and trained on a handful of publicly available datasets, and their performance does not trans-fer well to images captured with previously unseen cam-eras. Performing demosaicing for a specific MSFA requires training a new model, which is time-consuming and may require capturing new datasets, which hinders usability of SOTA models. We demonstrate, that demosaicing models with near SOTA performance can be trained using existing RGB datasets with simple hyperspectral augmentations. By performing random band reordering in the MSFA during training, our models seamlesly work with different MSFA. Conducted experiments show good quantitive and qualitative results. Our code will be made publicly available at github.com/ramuszl/cube-it.