Releasing the full artefacts behind an optical music recognition (OMR) benchmark, so the published figures can be recomputed rather than trusted.<\/p>\n
**Contents**
– Ground-truth MusicXML for 31 engraved pages from public-domain sources
– The same 31 pages after a deterministic phone-photograph degradation
– Raw MusicXML output from three engines on both suites: a commercial pipeline, Audiveris 5.11.0, oemer 0.1.x
– Transcriptions of 60 systems from OLiMPiC (ICDAR 2024) by two engines
– `score.py` \u2014 the scorer, Python standard library only, no dependencies
– `photo_sim.py` \u2014 the degradation model
– Per-piece results as JSON, plus the exact competitor CLI invocations<\/p>\n
**Metric:** strict F1 at note level \u2014 pitch, onset and duration must all match, with a 1\/64-note tolerance. Notes only; rests, clefs, key and time signatures, repeats and dynamics are not scored.<\/p>\n
**Size:** 2.4 MB. `python3 score.py .` reproduces the published table in about ten seconds.<\/p>\n
**Licence:** CC BY 4.0 for the engraved notation, transcriptions and scripts. The OLiMPiC-derived transcriptions inherit that dataset’s CC BY-SA; its ground truth is not redistributed.<\/p>\n
**Disclosure:** I work on one of the three engines. The scorer contains no engine-specific branching and a published erratum corrected a flaw that had been penalising a competitor.<\/p>\n
https:\/\/www.trysheetly.com\/compare\/omr-benchmark<\/a><\/p>\n<\/div>\n submitted by \/u\/alyimsa <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
