Head pose Estimation Using Convolutional Neural Networks

Head pose estimation is a fundamental problem in computer vision. Several methods has been proposed to solve this problem. Most existing methods use traditional computer vision methods and existing method of using neural networks works on depth bitmaps. In this project, we explore using convolutional neural networks (CNNs) that take RGB image as input to estimate the head pose. We use regression as the estimation approach. We explored the effect of different regularization strength and face alignment in our estimation. By using a CNN whose architecture is similar to VGG-nagadomi to train on IHDB head pose dataset, we can get a test regression euclidean loss of less than 0.0113, equivalent to average error of 20◦ of spherical distance, 4 times smaller than not using face alignment. We also proved that proper regularization strength could prevent overfitting thus reduce test loss.