Comparison of Facial Landmark Detection Methods for Micro-Expressions Analysis

In the paper the problem of the facial landmark detection is investigated. This task is considered as the stage of the pipeline for micro-expression analysis. The classification of the known facial landmark detection methods is represented. The most promising methods are selected for the experimental comparison. These methods are based on the cascaded regression and on the deep learning. The known open source software implementations for them are chosen, viz. the OpenFace library and the MediaPipe framework. The experiments are carried out on the images from the Spontaneous Actions and Micro-Movements (SAMM) dataset. To evaluate and compare the obtained results for the two tested methods the reference image is constructed manually. Average deviations of the detected landmarks from the referenced are estimated. The comparison shows that a little better precision of the facial landmark detection is achieved using the deep learning based method.