Open-sourced OMR benchmark that lets anyone recompute the numbers
A new optical music recognition benchmark ships more than claims: ground-truth MusicXML for 31 engraved pages, phone-degraded photocopies of the same scores, and raw output from three engines, a commercial pipeline, Audiveris 5.11.0, and oemer. The included scorer, score.py, needs zero dependencies, uses strict note-level F1, and rebuilds the published table in about ten seconds. Compact at 2.4 MB under CC BY, it turns OMR figures from trust-me into run-it-yourself.