Learn to Combine Multiple Hypotheses for Accurate Face Alignment

In this paper, we present the details of our method in attending the 300 Faces in-the-wild (300W) challenge. We build our method on cascade regression framework, where a series of regressors are utilized to progressively refine the shape initialized by face detector. In cascade regression, we use the HOG feature in a multi-scale manner, where the large pose validation is handled in early stages by HOG feature at large scale, and then shape is refined at later stages with HOG feature at small scale. We observe that the performance of the cascade regression method decreases when the initialization provided by face detector is not accurate enough (for faces with large appearance variations, face detection is still a challenging problem). To handle the problem, we propose to generate multiple hypotheses, and then learn to rank or combine these hypotheses to get the final result. The parameters in both learn to rank and learn to combine can be learned in a structural SVM framework. Despite the simplicity of our method, it achieves state-of-the-art performance on LFPW, and dramatically outperforms the baseline AAM on the 300-W challenge.