Cascaded Regression Based Face Alignment Algorithm

Face alignment has been one of the most active research areas in computer vision,which is also the key to many face analysis technologies,such as face recognition and face beautification.Recently,cascaded regression has proven to be one of the most successful algorithms for face alignment,which reaches the state-of-the-art performance both in accuracy and speed.Cascaded regression linearly combines several shape regressors to iteratively refine an initial shape to estimate the ground truth.Nowadays,most cascaded regression based methods focus on improving the regression ability of shape regressors,or extracting more effective geometry invariant face features.Despite these methods have achieved some progress on accuracy of face alignment,they ignore the quality of initial shapes.Initial shape plays an important role in cascaded regression,because algorithm would only obtain local optimal solution or even could not converge if the initial shape is far away from the ground truth.So,existing approaches are lack of capacity for dealing with some complex face alignment tasks as faces have extremely poses or exaggerate expression.To solve the above problem,we propose an initial shape estimation algorithm and a multipose cascaded regression framework.In this paper,we view initial shape estimation as a regression task,which regresses the ground truth shape from a holistic face feature by building a random regression forest model.Our algorithm is used to generate initial shapes in higher quality,so as to improve the accuracy.Meanwhile,we build the multi-pose cascaded regression framework using divide-and-conquer method on face pose.By doing this,different poses are trained independently,and every face can be regressed by a cascaded regressor after inferring its pose from the estimated initial shape.Experiments on some typical datasets show that our initial shape estimation algorithm generates initial shapes which are much more close to ground truth shapes compared with other initialization schemes,and it also decreases the error of existing cascaded regression based approaches by a large margin.Moreover,our multi-pose cascaded regression framework achieves comparatively or even more accurate results compared with other state-of-the-art face alignment algorithms.