Tinyposenet: A Fast And Compact Deep Network For Robust Head Pose Estimation

As an inherent attribute of human, head pose plays an important role in many tasks. In this paper, we formulate head pose estimation in different directions as a multi-task regression problem, and propose a fast, compact and robust head pose estimation model, named TinyPoseNet. Specifically, we combine the tasks of head pose estimation in different directions into one joint learning task and design the whole model based on the principle of "being deeper" and "being thinner" to obtain a tiny model with specially designed types and particular small numbers of filters. We perform thorough experiments on 3 types of test sets and compare our method with others from several different aspects, including the accuracy, the speed, the compactness and so on. In addition, we introduce large angle data in Multi-PIE to verify the ability of dealing with large-scale pose in practice. All the experiments demonstrate the advantages of the proposed model.