it's correct that we have "only" trained it on a couple of hundreds of classes, so it will only guess from that dataset.