Enhancing Human Pose Estimation in Ancient Vase Paintings via Perceptually-grounded Style Transfer Learning

Human pose estimation (HPE) is a central part of understanding the visual narration and body movements characters depicted in artwork collections, such as Greek vase paintings. Unfortunately, existing HPE methods do not generalise well across domains resulting poorly recognised poses. Therefore, we propose two step approach: (1) adapting dataset natural images known person annotations to style paintings by means image style-transfer. We introduce perceptually-grounded transfer training enforce perceptual consistency. Then, fine-tune base model with this newly created dataset. show that using style-transfer learning significantly improves SOTA performance on unlabelled data more than 6% mean average precision (mAP) recall (mAR) . (2) To improve already strong results further, small (ClassArch) consisting ancient from 6–5th century BCE annotations. fine-tuning style-transferred further. In thorough ablation study, give targeted analysis influence intensities, revealing learns generic domain styles. Additionally, provide pose-based retrieval demonstrate effectiveness our method. The code pretrained models can be found at https://github.com/angelvillar96/STLPose