Hybrid Deep Convolutional Network for Face Alignment and Head Pose Estimation

Zhiyong Wang, Jingjing Liu, Honghai Liu

Abstract: Face alignment has been an important focus of vision research because it is the most fundamental step in face analysis, reconstruction, and applications of emotion and attention. However, face alignment still suffers from some problems, such as lack of stability and poor performance in practical applications due to occlusion, illumination, and high training costs. This paper proposes a Dual-Task Hybrid Deep Convolutional Network (DHDCN) to estimate head pose and facial landmark locations simultaneously. By connecting the multi-level features, the local features and global features can be effectively fused. Features common to both tasks are learned in the initial stages of the network, and later stages will train the two tasks independently. Although the results have some gaps compared to the state-of-the-art results, it also demonstrates the feasibility and potential of learning both tasks simultaneously.