An Efficient Model for Simultaneous Face Detection, Pose Estimation and Landmark Localisation

A system that can perform multitasking including face Detection, pose Estimation and landmark Localisation (named DEL) is a core in many face recognition-based applications. Additionally, a wide range of applications that usually require machines to understand human face expressions can be obtained such us autonomous vehicles, patient observation or human-computer interaction (HCI). Typically, a DEL system is built by assembling three separate modules (face detection, pose estimation and landmark localisation) and performs them in succession. However, a recent approach of integrating three tasks into a single module shows a significant increase in performance. The main drawback of this approach is too time-consuming because of dozens of cross-correlation operations. In this paper, we propose an Efficient DEL (EDEL) model which is based on the integrated approach but more efficient. The key idea is to transform cross-correlations into matrix operations and apply matrix techniques to speed up the system. The experimental results on the AFW database reveal that the proposed model can perform 3 times faster in cross correlation time than the state-of-the-art integrated DELs but still remains the same level of accuracy.