[stubish]

One of the problems with these sorts of machine learning applications, including this exact use case elsewhere, is that they have been extremely sensitive to the imaging equipment used. Train it on a dataset of images from one source and it is only accurate on images from that source. Possibly only accurate on images from the exact same device. For home use, it needs a huge training set of images taken by all sorts of devices in all sorts of lighting conditions. And then the system will need to be improved until the error rates become useful.