Recently, several new approaches that do not use boosted cascade have been
proposed. Zhu et al. [32] used a mixture of deformable part models to capture
large face variations under different viewpoints and expressions. The model is
comprehensive and allows to estimate the head pose and facial points at the
same time in addition to detection. Shen et al. [24] proposed the first exemplarbased
face detector and exploited advanced image retrieval techniques to avoid
the expensive sliding window search. Both methods are better than Viola-Jones
style detectors on wild and challenging datasets [32, 9]. However, those accurate
detectors, as well as the state-of-the-art commercial face detector in Google
Picasa, are all quite slow due to their high complexity. Detection in an image
takes a few seconds and this makes such detectors unpractical for many scenarios.