REAL TIME FACIAL LANDMARK PREDICTION WITH REGRESSION TREES AND ONLINE MULTIPLE INSTANCE LEARNING

Abstract

Detection of Facial feature landmark is implemented by the proposed and efficient algorithm which works absolutely in real-time. Recent algorithms detecting facial features on 2D images requires good quality images. The method demonstrates how an Ensemble of Regression trees with Online Multiple Instance Learning (MIL), can be used for the estimation of face’s landmark positions, thus achieving actual real time speed and performance with the accurate predictions. Based on this, a general structure/framework is presented based on gradient boosting on learning the regression trees ensemble. This procedure optimizes the desired metric namely, sum of square error loss, along with instance label classifiers helping in achieving superior results with real-time performance. Thus significant improvement is achieved over state of the art methods. Lastly, we analyze the how appropriate priors helps with the selection of efficient features in the structure of image data.