Head Pose Estimation with Siamese Convolutional Neural Network
Fuxun Gao, Chaoli Wang · 2019 IEEE 8th Data Driven Control and Learning Systems Conference (DDCLS) · 2019
In order to get rid of the dependence on keypoints and improve the accuracy of pose estimation, in this paper, we propose a method of estimating the 3D head pose using a convolutional neural network with Siamese structure. Firstly, the rank labels of head pose which used to train the Siamese network can be automatically generated from the continuous head pose labels. The Siamese network can rank the head pose deflect levels and this step is equivalent to coarse classification. Secondly, after Siamese network is trained, the continuous raw pose labels were used to fine-tuning a branch of Siamese network and let the network regress the continuous pose ground truth. For avoiding duplicate computation caused by Siamese network, we add a ensemble layer to the network. In addition, high intensity brightness adjustment and Gaussian blur are imposed on images to distort images in data augmentation, so our method will achieve perfect performances in low quality images. Experiments show that our method has higher accuracy than the state-of-the-art methods of estimating head pose from RGB images, stronger robustness than the method of head pose estimation with keypoints, and wider application range than the method of head pose estimation with depth data.