Deep head pose estimation for faces in the wile and its transfer learning
Bao Hanh Tran, Yong-Guk Kim · 2017
Three-dimensional head pose estimation has been an important and challenging task in computer vision partly because of its diverse applications. In this paper, we propose a new method to estimate head pose for the faces in the wild using deep neural network based on classification, rather than conventional regression. The network consists of three CNNs, corresponding to three head pose components, i.e. yaw, pitch and roll, respectively, and each CNN has 3 channels for color input. We have trained the network using the representative face datasets in the wild, such as AFLW, AFW and YouTube, and found that it outperforms the state-of-art algorithms with a substantial margin. We have also explored transfer learning among them: training the network with a dataset and testing it against others. It is found that there is directional difference between them. And yet, result obtained by transferring from AFLW to others is still better than those from recent studies.