cobernhart / retrofaceid

RetroFaceID: Face Recognition in Historical Image Archives

The Theatre Museum Vienna possesses an extensive archive of approximately 1.6 million historical images, a significant portion of which remains unlabeled. This paper introduces RetroFaceID, a state-of-the-art face recognition tool designed to improve the labeling process for historians. Built on advanced machine learning algorithms, the tool offers a semi-automated workflow that significantly enhances productivity and label quality. To validate its effectiveness, a new dataset is created from the Theatre Museum's archive, incorporating multiple types of intra-class variations such as age, pose, and facial expressions. The tool's face recognition model achieves an accuracy of 79.6% on this dataset, outperforming existing approaches. The results demonstrate the tool's transformative potential for historical research, enabling efficient and objective search, identification, and verification of historical figures in large-scale datasets.