mAP
YOLOv11S reached 0.4126 mAP@[0.5:0.95] on DeepScoresV2 Dense.
Evaluation
Detection quality and reconstruction quality are reported separately.
PyClef separates object-detection metrics from functional pitch reconstruction metrics, since high mAP does not necessarily imply correct musical reconstruction.
TRA
With YOLOv11S, MIRP increased TRA from 0.04 to 0.975.
Model trade-off
MIRP reduces the practical gap between model sizes.
YOLOv11L reached 0.51379 mAP, but both YOLOv11S and YOLOv11L exceeded 97% TRA after MIRP.
Summary
Reported comparison points.
| Metric | YOLOv11s | YOLOv11l |
|---|---|---|
| mAP@[0.5:0.95] | 0.4126 | 0.51379 |
| Pitch Reconstruction Rate | 0.04 -> 0.975 | 0.08 -> 0.971 |
| Model size | Smaller | Larger |
| Training cost | Lower | Higher |