AI Based Actors Identification with High Intra-Class Variations

Abstract

While deep learning based face recognition sur-passes human performance in constrained settings, it still strug-gles to achieve similar results applied in completely unconstrained settings. This paper explores the effectiveness of state-of-the-art face recognition models in the specific case of identifying actors in a historical photography collection of a Theatre Museum. Actors can be pictured at different angles and poses, at a different age, with masks and costumes leading to strong intra-class variations. In addition, images might show signs of decay due to their historical nature, further increasing the difficulty for a face recognition model to make correct predictions. This paper shows that ElasticFace, a face recognition model trained using a novel learning loss strategy, achieves 79.6% accuracy on the museum's photo database. Based on those outcomes, deploying face recognition to analyse historical image collections delivers valuable results for historians.