Melodious
From printed sheet music to editable, playable digital scores.
Melodious turns a photograph of printed sheet music into a full musical document: 136-class detection, graph relationship inference, inspectable confidence, then editable MusicXML, playable MIDI, engraved notation, and downloadable artifacts. On held-out DeepScoresV2 testing it reached 0.852 mAP@0.5, beating published Faster R-CNN + HRNet (0.799) and Deep Watershed (0.503).
Real product surfaces from the Melodious demo and repository.
View source on GitHub Open live productDeepScoresV2 · held-out detector test
Beat published benchmarks on mAP@0.5
0.852mAP@0.5
Melodious outperformed published DeepScoresV2 results on the same metric, including Faster R-CNN + HRNet at 0.799 and Deep Watershed Detector at 0.503.
- Melodious0.852Melodious
- Faster R-CNN + HRNet0.799Published benchmark
- Deep Watershed Detector0.503Published benchmark
0.707held-out test mAP@0.5:0.95
0.839held-out test F1@0.5
22%relative graph-model uplift over deterministic geometry on the same 14 pages and 48,174 edges
