Real-Time Ear Landmark Detection Using Ensemble of Regression Trees

Abstract

Human face landmark detection algorithms have numerous applications. Current face landmark detection algorithms limit themselves to features around eyes, nose, cheeks and lips. Face landmark detection combined with augmented reality technology has given rise to commercially popular virtual try-on applications. To realize use cases of virtual jewelry try-on like earrings on smartphones, landmark points of human ear is required, but this field is not much explored in the literature. Existing methods are not accurate enough in different face poses and lighting conditions. Proposed method offers solution for ear landmark detection considering the computational requirements of mobility devices, and comprises ear localization followed by ear landmark detection. It adopts Haar cascade based model for ear localization and an Ensemble of Regression Trees for ear landmark detection. The experimental results and comparison with state-of-the-art methods show that the proposed method accurately localizes the ear, provides correct landmark points and is fast enough to run on mobility devices with low memory footprints. Comparison with popular methods shows the novelty points in the proposed approach.