Facial feature localization using MOSSE correlation filters

Accurately measuring the location of facial features is an important step in many face recognition algorithms. Every face is unique which means localization needs to be tolerant of differences between individual subjects. Additionally, changing illumination, poor focus, and deformation due to expression changes complicate the problem. This paper introduces a method for locating facial features that uses Minimum Output Sum of Squared Error (MOSSE) correlation filters to model object appearance and is combined with a Robust Active Shape Model (ASM) to model facial geometry. It is demonstrated that MOSSE correlation filters outperform Stasm (an open source ASM implementation), Gabor Jets and in some cases even matches human performance.