Gotsuzuri maker finds errors in own GT labels before AI training

The record picks up from an earlier prototype of Gotsuzuri, an Android app that turns photos of a real Go board into game records. Even with the four corners aligned, shadows, reflections, and the height of the stones remained obstacles to recognizing stones, and a later attempt to reduce misassignment of neighboring stones to the wrong intersection by jointly mapping stone candidates to intersections did not pan out: he reports that it improved some cases but made the board worse overall, so the existing judgment was kept.

The emphasis then shifted from recognition logic to the data itself. To compare methods across different stone layouts, he added 6 setups and 12 photos, classifying each intersection as empty, black, white, or undiscernible, and built tools to reuse shooting layout diagrams and already-checked ground truth as drafts. Reviewing all 21 images with a side-by-side photo and top-down GT diagram surfaced his own labeling slips, and on the upper edge some black stones could not be pinned to an intersection by a human eye. Rather than fill those in as empty to make the counts tidy, he left them as undiscernible, and he notes that the 21 images include multiple angles of the same setups and are not 21 independent boards, with the home board and room still a source of bias.

The next planned step is a small training experiment: classify empty, black, and white from images around each intersection, first with images alone and then with grid spacing and tilt added, using PyTorch and AWS SageMaker. This is preparation rather than a result, since no AI model has been trained yet.