Eye feature points detection by CNN with strict geometric constraint

The detection accuracy of facial landmarks or eye feature points is influenced by geometric constraint between the points. However, this constraint is far from being research in existing convolutional neural network (CNN) based points detection. Whether strict geometric constraint can improve the performance is not studied yet. In this paper, we propose a new approach to estimate the eye feature points by using single CNN. A deep network containing three convolutional layers is built for points detection. To analyze the influence of geometric constraint on CNN based points detection, three definitions of the eye feature points are proposed and used for calibration. The experiments show that excellent performance is achieved by our method, which prove the importance of the strict geometric constraint in points detection based on CNN. In addition, the proposed method achieves high accuracy of 96.0% at 5% detection error, but need less computing time than the cascade structure.