Face alignment using structured random regressors combined with statistical shape model fitting

Face alignment involves locating several facial parts such as eyes, nose and mouth, and has been popularly tackled by fitting deformable models. In this paper, we explore the effect of the combination of structured random regressors and Constrained Local Models (CLMs). Unlike most previous CLMs, we proposed a novel structured random regressors to give a joint prediction rather than pursuing independence while learning the response map for each facial part. In our method, we first present a fast algorithm to learn local graph, which will then be efficiently incorporated into the random regressors. Finally we regularize the output using a global shape model. The benefits of our method are: (i) random regressors allow integration of votes from nearby regions, which can handle various appearance variations, (ii) local graph encodes local geometry and enables joint learning of features of facial parts, (iii) the global model regularizes the result to ensure a plausible final shape. Experimentally, we found our methods to converge easily. We conjecture that structured random regressors can efficiently select good candidate points. Encouraging experimental results are obtained on several publicly available face databases.