F3ED Can't Call an Ace: Fixing a NeurIPS 2024 Tennis Model

I built a tennis broadcast pipeline this spring and ended up running F3ED, the NeurIPS 2024 shot detector, on a couple of ATP Challenger matches. F3ED is a good model. It also kept labeling clear aces as “unforced errors”, which is what this post is about. Code: github.com/philippdubach/tennis-vision. F3ED (NeurIPS 2024) detects shots well. The catch is the outcome head, which has 4 classes: in, winner, forced-err, unforced-err. There’s no class for ace, double_fault, or first_serve_fault. Those events aren’t shot properties; they’re score-grammar, and they need state from outside the shot itself. I audited 11 single-shot serve rallies F3ED labeled unforced-err. 7 are first-serve-faults. 1 is an ace. Only 3 are genuine unforced-errors. 73% mislabeled by tennis’s own definition. The fix is a 30-line reconciler that reads the scoreboard. OCR isn’t novel here. What I haven’t seen anyone do is plug it back into runtime label correction, which is what makes the difference. N=44 rallies…