Abstract
This paper addresses the problem of Face Alignment fora single image. We show how an ensemble of regressiontrees can be used to estimate the face’s landmark positionsdirectly from a sparse subset of pixel intensities, achievingsuper-realtime performance with high quality predictions.We present a general framework based on gradient boostingfor learning an ensemble of regression trees that optimizesthe sum of square error loss and naturally handles missingor partially labelled data. We show how using appropriatepriors exploiting the structure of image data helps with ef-ficient feature selection. Different regularization strategiesand its importance to combat overfitting are also investi-gated. In addition, we analyse the effect of the quantity oftraining data on the accuracy of the predictions and explorethe effect of data augmentation using synthesized data.