Pose Annotation Project for Artworks: A participatory annotation platform for automated body pose estimation in art

Many projects in the domain of digital humanities have shown a growing interest for object recognition in artworks [1], and, more specifically, for the task of Human Pose Estimation (HPE). HPE is a well defined challenge in computer vision and benefits from large training datasets of annotated images that allow the creation of highly performative machine learning models. However, these datasets are usually made of recent photographs collected in the web, capturing a loose western perspective with all its biases Moreover, their content diverges from artworks created in past Centuries. For the case of figurative paintings, visual features, such as brush marks and color shades, but also the morphology of bodies, dressing codes or the overall composition, have an impact on the accuracy of HPE models. These many differences do not enable the field of digital art history to fully benefit from the performance of existing models and the lack of a proper training dataset for artworks makes it difficult to retrain them. Therefore, we present the Pose Annotation Project for Artworks (PAPA), which goal is to produce a training dataset for body pose estimation on artworks through a platform for participatory annotations of images.