A Novel Facial Feature Localization Method Using Probabilistic-like Output

Object detection technique generally could not localize facial features precisely because there is a tradeoff between detection robustness and facial feature support [2]. To address this problem, we propose a method to calculate probabilistic-like output for each pixel of image. This output describes the similarity of a patch of image to the training samples. The probabilistic-like output is afterwards used to locate the feature points using a localization approach we proposed. Our algorithm for facial feature localization is fast, accurate and robust. It takes only about 10ms on a computer with P4 CPU to locate five feature points including eye centers, nose tip and mouth corners. The localization accuracy is comparable with hand labeled results. And experimental results on a large size of face database (more than 12,000 images from PIE [10]) demonstrate that the proposed method is robust to the variances of pose, illumination, expression and other appearance factors, such as glasses and beard.