In this work, we focus our attention on the similarity among works of art based on human poses and the actions they represent, moving from the concept of Pathosformel in Aby Warburg. This form of similarity is investigated by performing a pose clustering of the human poses, which are modeled as 2D skeletons and are defined as sets of 14 points connected by limbs. To build a dataset of properly annotated artwork images (that is, including the 2D skeletons of the human figures represented), we relied on one of the most popular, recent, and accurate deep learning frameworks for pose tracking of human figures, namely OpenPose. To measure the similarity between human poses, two alternative distance functions are proposed. Moreover, we developed a modified version of the K-Medians algorithm to cluster similar poses and to find a limited number of poses that are representative of the whole dataset. The proposed approach was also compared to two popular clustering strategies, that is, K-Means and the Nearest Point Algorithm, showing higher robustness to outliers. Finally, we assessed the validity of the proposed framework, which we named POSE-ID-on, in both a qualitative and in a quantitative way by simulating a supervised setting, since we lacked a proper reference for comparison.

中文翻译：

---

POSE-ID-on-一种用于艺术品姿势聚类的新颖框架

在这项工作中，我们将注意力集中在基于人体姿势及其所代表的动作的艺术品之间的相似性，而这是从Pathosformel的概念出发的在Aby Warburg。通过对人类姿势进行姿势聚类来研究这种相似性形式，这些姿势被建模为2D骨架，并定义为由肢体连接的14个点的集合。为了构建带有正确注释的艺术品图像（即包括所表示的人物的2D骨架）的数据集，我们依赖于最流行，最新且最准确的深度学习框架之一，即人物姿态跟踪，即OpenPose。为了测量人体姿势之间的相似性，提出了两个替代的距离函数。此外，我们开发了K-Medians算法的修改版本，以聚类相似的姿势，并找到代表整个数据集的有限数量的姿势。还将该提议的方法与两种流行的聚类策略进行了比较，即 K均值和最近点算法，对异常值显示出更高的鲁棒性。最后，由于缺乏适当的比较参考，我们通过模拟监督环境以定性和定量方式评估了拟议框架（我们命名为POSE-ID-on）的有效性。