Few-Shot Object Detection: Application to Medieval Musicological Studies.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Humanities

Background:

• Few-shot object detection is challenging for cultural heritage datasets due to style variations.
• Existing methods struggle with detecting small objects in non-standard image styles.

Purpose of the Study:

• To present a simple, effective black-box method for few-shot object detection.
• To introduce the Medieval Musicological Studies Dataset (MMSD) for challenging object detection tasks.
• To benchmark the proposed method against state-of-the-art object detection models.

Main Methods:

• Developed a novel black-box few-shot object detection approach compatible with various models.
• Created and annotated the MMSD dataset with 693 samples across five classes.
• Evaluated the method on YOLOv4, (Mask/Faster) R-CNN, and ViT/Swin-t using two benchmarking strategies.

Main Results:

• The proposed method consistently improved object detection performance compared to traditional transfer learning.
• Performance gains were observed across different model architectures (YOLOv4, R-CNN, ViT/Swin-t).
• The MMSD dataset proved more challenging than existing benchmarks due to style diversity.

Conclusions:

• The presented method offers a robust solution for few-shot object detection in specialized domains like cultural heritage.
• The MMSD dataset provides a valuable resource for advancing research in this area.
• The approach demonstrates significant improvements over standard transfer learning techniques.