Good work! Training on vector pictures instead of rasterised images seems such a good way to go. With some related data, I imagine this can also be colored.