An Effective Initialization for ASM-Based Methods

Locating facial feature points is an important step for many facial image analysis tasks. Over the past few years, Active Shape Model (ASM) has become one of the most popular approaches to solve this problem. However, ASM-based methods are sensitive to initialization errors caused by poor face detection results. In this paper, an effective initialization for the ASM-based methods is proposed by our improved initial 8-landmarks ASM model together with Viola-Jones eye detection. In particular, we apply the 8-landmarks ASM model with the position of eyes from eye detector as reference points. After that, we choose several 76-landmark candidates from the training set that have the key feature points related to the result of the previous 8-landmark model. The best candidate has the lowest fitting errors with the test image and is used as initialization. To evaluate the performance of our work, we conduct the experiments on the MUCT and LFPW database. Compared to the latest ASM implementation, MASM, our proposed can improved the MASM by an average accuracy of 44% when dealing with poor initialization.