685
SMB Samples
Comparison Setup
Dataset
SMB with 685 sheet music images across multiple texture categories and ground truth in kern format.
Models
Audiveris, Audiveris-Scaled, HOMR, and OeMeR. Audiveris-Scaled is Audiveris with pre-processing upscaling.
Evaluation
OMR-NED as a distance metric between prediction and ground truth, complemented by error-type analysis.
Key Findings
Four-model comparison (n=104)
Median OMR-NED: HOMR 0.243, Audiveris-Scaled 0.250, Audiveris 0.321, OeMeR 0.715.
HOMR leads, while OeMeR clearly trails.
Three-model comparison (n=453)
With more challenging Pianoform samples, all values increase: HOMR 0.603, Audiveris-Scaled 0.660, OeMeR 0.753.
Performance ranking remains stable: HOMR > Audiveris-Scaled > OeMeR.
Coverage
Successfully processed SMB samples: Audiveris 29.3%, Audiveris-Scaled 78.5%, HOMR 99.3%, OeMeR 84.8%.
Upscaling makes Audiveris usable on significantly more pages.
Result Plots
4 Models (n=104)
3 Models (n=453)
Conclusion
HOMR delivers the best overall benchmark performance. Audiveris-Scaled is a practical compromise with significantly higher coverage than Audiveris. In this setup, OeMeR shows consistently weaker results.
Key takeaway for future work: error profiles are model-specific, so improvements should be designed per model rather than with a single global fix.