Assemblies of Points: Strategies to Art-historical Human Pose Estimation and Retrieval

This paper attempts to construct a virtual space of possibilities for the historical embedding of the human figure, and its posture, in the visual arts by proposing a view-invariant approach to Human Pose Retrieval (HPR) that resolves the ambiguity of projecting three-dimensional postures onto their two-dimensional counterparts. In addition, we present a refined approach for classifying human postures using a support set of 110 art-historical reference postures. The method’s effectiveness on art-historical images was validated through a two-stage approach of broad-scale filtering preceded by a detailed examination of individual postures: an aggregate-level analysis of metadata-induced hotspots, and an individual-level analysis of topic-centered query postures. As a case study, we examined depictions of the crucified, which often adhere to a canonical form with little variation over time — making it an ideal subject for testing the validity of Deep Learning (DL)-based methods.