Dual Training for Transfer Learning: Application on Medieval Studies

Abstract

Object detection and Image classification have witnessed tremendous improvement in fixing domain gaps between training and deployment data. However, Transfer Learning is still the best method for Object Detection that provides resilient results. This paper presents two main contributions. First, we developed a benchmark annotated dataset of stringed musical instruments in artworks from the Middle Ages (827 images), the Early Modern Age (165 images), and the Contemporary History (10258 images). Second, we present a new Transfer Learning method for Object Detection models that is non-intrusive, simple, reproducible, and model-independent. Our method iteratively trains the black box Object detection models on the source and target datasets and shifts the focus between them dynamically to improve the results on the target dataset while maintaining the performance on the source dataset. Our method was thoroughly evaluated against several Transfer Learning methods on YOLOv4, YOLOv5, PP-YOLO, and Detectron2 with their respective versions. The experimental results show that our method outperforms existing Transfer Learning techniques with over 8.83% F1-score on our dataset.