Learning Similarity for Discovering Inspirations of Western Arts in Japanese Culture.

Several paintings by Japanese artists in the beginning of 20th century were largely inspired by works of western artists. Finding correspondences between the Japanese and the western artworks can reveal how the western arts were introduced in Japan. Until now, to discover such correspondences, art historians usually annotated them manually. This is a tedious process which generally requires a lot of effort and time. In computer vision literature, there are several techniques that can find similarities in images. To find such similarities some techniques are based on objects appearing in the images, while some techniques compare fine-grain details. However, inspirations in art illustrations are sometimes from global outlines of the images. Another difficulty is that annotations of correspondences are rare in historical data. This makes a lot of techniques which are based on supervised learning not applicable. In this paper, we propose a novel technique to find correspondences between two related artworks, which compares the global outlines information. It is based on Siamese neural networks (SNNs) and self-supervised learning method. In addition, we create a dataset of illustrations from two different types of artworks: one from Japanese artists, Seiki Kuroda, and one from western artists, Raphaël Collin. Correspondence annotations are also given. We evaluate the algorithm using recall@k as metrics, and also qualitatively show that the proposed method provides profiles of image correspondences different from the state of the art.