Efficient Construction of Face Image Dataset for Head Pose Estimation

The purpose of this study is to estimate the head pose of a person in an operating room. Most publicly available datasets for head pose estimation (HPE) are images taken from the front, making it difficult to estimate the head pose of the back of the head. In addition, since they are not intended to healthcare professional's head pose, they do not include scenes in which a person is wearing a medical cap and a mask. Therefore, existing datasets do not match the purpose of our project and cannot correctly estimate head pose. Therefore, in this paper, we construct our original dataset. Since constructing a dataset is labor-intensive, we propose a method to construct a dataset in which data is distributed efficiently and uniformly using a turntable. A dataset consisting of approximately 20,000 images was constructed using the proposed method. We evaluated the accuracy of HPE using an existing deep learning model and obtained MAE of 5.67° for the pitch angle and 6.37° for the yaw angle. Furthermore, we compared the results with models trained on other datasets for an actual operating room scene and showed that our constructed dataset is useful.