Deep learning-based human pose estimation towards artworks classification

Abstract

Human pose detection has attracted more attention in recent years, particularly with various applications such as human-computer interaction, motion recognition, action prediction, gaming, sign language translation, video surveillance and human tracking. For example, it is challenging to analyse the proportions of the human body in historical artwork collections for classifying genres, styles, and artists. Unfortunately, most of existing detection methods do not generalize well across artworks, resulting in poorly recognized differences in the proportions. Therefore, we present a large-scale analysis of 130,000+ paintings and ninety-nine human pose estimation (HPE) methods to show that different artistic styles have a distinct average degree of human proportions. We have further verified that the Topdown Heatmap + Scnet algorithm with a threshold of 0.3 can classify artworks effectively and fully distinguish historical epochs. This analysis is a baseline for researchers to discover new techniques towards finding the canon of art based on symmetry and proportions.