[P] Magic Cards classifier in streaming

My master degree's thesis consists in a tool that detects and classifies Magic: the Gathering cards during a tabletop streaming (no online games). MtG official tournaments have some streaming procedures to follow, like a top-view camera, card positioning on the table, a certain lighting etc that can help the machine a lot by avoiding harsh conditions. Main Issue: There are around 20.000 different cards, and some of them have more than 1 image representing em (alternative artworks or frames). Is classification with that number of classes even doable in a reasonable training time? How to react: I could accept training on a smaller subset (like maybe the newest cards), but I would like to create something useful, and not just a demonstrative project.