Experimenting active and sequential learning in a medieval music manuscript

Sachin Sharma, Federico Simonetta, Michele Flammini

Introduction

Optical Music Recognition (OMR) serves as the "OCR for music," yet it encounters significant obstacles when applied to historical manuscripts. Unlike modern printed scores, medieval manuscripts—such as the 13th-century Italian Lauda—are hand-drawn, physically degraded, and utilize non-standard notation. The primary bottleneck in digitizing these works is data scarcity; training robust deep learning models requires thousands of expert-level annotations. This study investigates whether Active Learning (AL)—where the model identifies samples it is most uncertain about for labeling—can reduce manual labor compared to Sequential Learning (SL), which follows the natural page-by-page order of a manuscript.

Methodology

The researchers implemented an object detection pipeline using the YOLOv8n architecture. The training followed an iterative cycle designed to compare the efficiency of different sampling strategies.

The iterative Active Learning cycle used to train the object detection model.

The process involved:

  1. Initial Training: Starting with a single annotated image.
  2. Inference: Running the model on the remaining unlabeled pool.
  3. Selection:
  4. AL Strategy: Selecting the 15 images with the lowest prediction confidence (highest uncertainty).
  5. SL Strategy: Selecting the next 15 images in the manuscript's physical sequence.
  6. Retraining: Updating the labeled dataset and repeating for 20 rounds.

The model was tasked with identifying nine symbol categories, including neumes (musical notes), clefs, staves, and text lines.

Dataset: The Cortonese Manuscript

The study utilizes a novel dataset derived from the I-Ct 91 manuscript (the "Cortonese"), a medieval source spanning the 12th to 16th centuries. The dataset consists of 340 images containing 7,015 bounding box annotations. The manuscript is characterized by multiple scribal hands and "non-mensural" notation, presenting a higher complexity than standard Western music notation.

Results and Discussion

The experimental results revealed a surprising "uncertainty trap" in the early stages of training. Sequential Learning initially outperformed Active Learning because the early pages of the manuscript happened to match the dominant scribal style in the test set. In contrast, AL's uncertainty-based sampling initially selected varied and rare regions, leading to a slower start.

Comparison of model performance metrics over rounds for (a) Active Learning and (b) Sequential Learning.

By rounds 8–10 (approximately 120–150 images), the performance gap closed. Once the model encountered sufficient variety, AL’s targeted selection allowed it to generalize effectively. By the final round (272 images), both strategies achieved nearly identical results:

  • AL: 77.3% mAP@50
  • SL: 77.4% mAP@50

Notably, both iterative strategies eventually outperformed a fully supervised baseline trained on the entire dataset at once. This suggests that the increased number of optimization steps in the iterative process contributes to superior final performance.

Conclusion

While Active Learning is theoretically more efficient, this study highlights that simple sequential annotation can be highly effective for heterogeneous historical data, especially when early samples are representative of the broader dataset. Future work will explore advanced AL strategies that balance uncertainty with representativeness and extend the framework to handle complex multi-page layout analysis and domain adaptation across different manuscripts.

Summary updated

BibTeX
@inproceedings{sharma2025experimenting,
  author = {Sachin Sharma and Federico Simonetta and Michele Flammini},
  title = {{Experimenting active and sequential learning in a medieval music manuscript}},
  booktitle = {MLSP},
  year = {2025},
  doi = {10.1109/mlsp62443.2025.11204321},
}
Figure