Semi-supervised Human Pose Estimation in Art-historical Images

Springstein, Matthias ; Schneider, Stefanie ; Althaus, Christian und Ewerth, Ralph (2022): Semi-supervised Human Pose Estimation in Art-historical Images. MM: Conference on Multimedia 2022, Lisboa, Portugal, 10. - 14. Oktober 2022. Magalhães, João (ed.) : In: Proceedings of the 30th ACM International Conference on Multimedia, New York, USA: Association for Computing Machinery. pp. 1107-1116

Abstract

Gesture as ‘language’ of non-verbal communication has been theo-retically established since the 17th century. However, its relevancefor the visual arts has been expressed only sporadically. This maybe primarily due to the sheer overwhelming amount of data thattraditionally had to be processed by hand. With the steady progressof digitization, though, a growing number of historical artifactshave been indexed and made available to the public, creating a needfor automatic retrieval of art-historical motifs with similar bodyconstellations or poses. Since the domain of art differs significantlyfrom existing real-world data sets for human pose estimation dueto its style variance, this presents new challenges. In this paper, wepropose a novel approach to estimate human poses in art-historicalimages. In contrast to previous work that attempts to bridge thedomain gap with pre-trained models or through style transfer, wesuggest semi-supervised learning for both object and keypoint de-tection. Furthermore, we introduce a novel domain-specific art dataset that includes both bounding box and keypoint annotations ofhuman figures. Our approach achieves significantly better resultsthan methods that use pre-trained models or style transfer.

• Item Type: Conference or Workshop Item (Paper)
• Faculties: History and Art History > Department of Art History > Art History
• Subjects: 700 Arts and recreation > 700 Arts
• ISBN: 9781450392037
• Place of Publication: New York, USA
• Language: English
• DFG: Gefördert durch die Deutsche Forschungsgemeinschaft (DFG) - 415796915