Annotating each move manually seems a lot faster than the MVP app in the article, which seems to take some 10 seconds to recognize the whole game state. It als

Annotating each move manually seems a lot faster than the MVP app in the article, which seems to take some 10 seconds to recognize the whole game state. It also seems suited towards an altogether different use case, where you don't already know the previous game state or the moves that resulted in it. This recognizes the entire game state over and over again, which is a lot of redundant work if you already know the previous game state and only consider valid moves from that state. For the purpose of recording an entire game, a much simpler computer vision approach would be possible based on the previous state. Just find the empty square where there was previously a piece, consider where it might have ended up according to its possible moves and look for changes in those squares. The hardest CV issue then is figuring out what a pawn was promoted to.