A comparative Study of Automatic Facial Landmark Detection

Abstract

Facial signs are associated with people’s health and general fitness. Among different facial signs, the facial landmark is one of the essential appearances of facial characters, which can be linked with people’s emotions, state of consciousness and health. Facial landmark detection can be used for recognising people’s expressions, monitoring the conscious status of people’s faces, or diagnosing neurological diseases. Recent advances in imaging technology and ever increasing computing power have opened up the possibility of automatic facial landmark analysis and assessment. Facial landmark detection algorithms play an important role in facial analysis tasks, such as expression recognition, face swapping and medical auxiliary diagnosis. As a result, the accuracy of the facial landmark localisation directly impacts the reliability of facial landmark based tasks .

The purpose of this project is to conduct a comparative study of existing vision based methodologies for detecting facial landmarks and to identify appropriate ones that could overcome the challenges, such as pose variation and exaggerated expression. Three effective facial landmark detection methodologies were selected and implemented in this project, including Deep Convolutional Neural Network (DCNN) Cascade, Deep Alignment Network (DAN), and Stacked Dense U-nets (SDU). In order to provide a thorough evaluation, three publicly available datasets were used for the benchmarking, such as Multi PIE, 300W and Menpo, which contain a large number of facial images with the variations in illumination, pose and expression as well as the occlusion. Through the evaluation based on different datasets, SDU was considered to have the best performance and it was adopted and implemented into a real time facial analysis system that contains landmark detection and assessment