Facial landmark localization via boosted and adaptive filters

Recently, the ASEF and MOSSE filters have exhibited impressive performance in facial keypoint localization, which is often a vital step in facial image analysis. Correlation outputs of training samples are first designed as Gaussians, and the filters are reversely constructed via averaging and summed error minimization in Fourier domain, where correlation can be efficiently computed by element-wise multiplication. To further improve the performance, this paper proposes two kinds of techniques extended from ASEF and MOSSE: (1) add an error correction module, and increase the weights of inaccurately detected samples under the framework of boosting algorithm; (2) iteratively adjust the synthetic outputs to be more adaptive to image contents. Experimental results show the proposed methods are superior to ASEF and MOSSE in keypoint finding.