Iterative adaptive synthetic correlation output filters

The average of synthetic exact filter (ASEF) and the minimum output sum of squared error (MOSSE) are two state-of-the-art correlation filters. An iterative adaptive method to boost their performance of target finding is proposed. ASEF and MOSSE stiffly assign distance-based Gaussians to the training images as synthetic correlation outputs, which drop the intensity information and may distort the filter. To alleviate it, synthetic outputs are iteratively adjusted under two considerations: (i) the correlation peak should locate at the target by giving more prominence to the target while suppressing other local maxima that are likely to be wrongly detected, and (ii) correlation values at different pixels should match the intensity context. Accordingly, the filter is updated. Comparative experiments in facial landmark localisation show the superiority of the proposed method over ASEF and MOSSE.