Also, a 10% error rate wouldn't be a big deal as long as it was easy to correct the position within a few seconds.
Indeed, the checkerboard itself gives valuable information about the pose of the checkerboard, it is even used for calibration in multi-view geometry: https://en.wikipedia.org/wiki/Chessboard_detection
Yes but does real world photos ever have that poor contrast? IMO, top down photo is worth exploring.
Compared to all the other feats of machine learning that have blown my mind, parsing a photo of a limited set of a handful of different piece variants, in two colors, that located on a grid, doesn't seem too difficult.