MIDL Vision 2022

Domain adaptation through anatomical constraints for 3d human pose estimation under the cover Alexander Bigalke is a PhD student at the University of Lübeck under the supervision of Mattias P. Heinrich. His work on 3D human pose estimation applied to patient monitoring has been picked for an oral presentation today, and he is here to tell us more. Deep learning dominates the state of the art for general human pose estimation, but there are still some challenges in the clinical setting. This paper explores the problem of monitoring patients who are lying in bed . One of the greatest difficulties in this situation is occlusion by blankets , which renders it impossible to create manual annotations and hampers the ability to train neural networks. “ We propose to overcome this problem by domain adaptation, ” Alexander tells us. “ We want to adapt the pose estimator from patients that are not occluded by blankets to patients that are occluded by blankets. ” 12 Oral Presentation VISION MIDL