Faster Real-Time Face Alignment Method on CPU

Face alignment for facial images captured in-the-wild is a challenging and important problem. In this work, we introduced a two-stage face alignment method in order to solve the problem of the normal face alignment method running slowly on the CPU. Using the residual error between ground truth and mean shape as a training label makes the network easier to converge. The joint input of heatmap and original data in the second stage deepens the feature learning of these landmarks, making the minimal network also has suitable performance. The convolution and pooling structure allow the network to be faster and have good learning ability. The test results on open datasets show that our method has a significant improvement in processing performance with real-time CPU speed of 1100 fps while maintaining high accuracy.