If you're going the ML way, I recommend an end-to-end approach to get significantly more performance (at a data preparation and training cost). Simply generate

If you're going the ML way, I recommend an end-to-end approach to get significantly more performance (at a data preparation and training cost). Simply generate a dataset of chess boards, and then render them using a 3D software of choice (pretty easy job in blender/etc.). For robustness then you can add various kinds of perturbations, like different camera angles, small displacements for each piece, etc.