Unified convolutional neural network for direct facial keypoints detection

We propose a novel approach to directly estimate the position of the facial keypoints via convolutional neural networks (CNN). Our method estimates the global position and the local positions from a unified CNN and combines them through a simplified optimization process. There are twofolds of advantages for our approach. First, the global geometrical position and the local detailed position of the facial keypoints are combined complementarily to avoid local minimums caused by occlusions and pose variations. Second, unlike the traditional method such as a cascade of multiple CNN, we propose a unified deep and large architecture network consisted by global position network and local position network. Our design shares most of computations for facial features between networks, and this efficient high-level features improves largely to the precise estimate of facial keypoints. We conduct comparative experiments with the state-of-the-art researches and commercial services. In experiments, our approach shows a remarkable performance.