A Data-Driven Approach to Improve 3D Head-Pose Estimation

Head-pose estimation from images is an important research topic in computer vision. Its many applications include detecting focus of attention, tracking driver behavior, and human-computer interaction. Recent research on head-pose estimation has focused on developing models based on deep convolutional neural networks (CNNs). These models are trained using transfer-learning and image augmentation to achieve better initiation states and robustness against occlusions. However, methods that use transfer-learning networks are usually aimed at general image recognition and offer no in-depth study of transfer learning from more task-related networks. Additionally, for the head-pose estimation, robustness against heavy occlusion, and noise such as motion blur and low-brightness are vital. In this paper, we propose a new image-augmentation approach that significantly improves the estimation accuracy of the head-pose model. We also propose a task-related weight initialization to further improve the estimation accuracy by studying internal activations of models trained for face-related tasks such as face-recognition. We test our head-pose estimation model on three challenging test sets and achieve better results to state-of-the-art methods.