Robustness of head pose estimation based on Gabor wavelets and graph matching

Gabor wavelet filter responses used as graph nodes in labelled graph matching have previously been shown to give high accuracy for head pose estimation from a single two-dimensional image. However, the robustness of this approach to additive or convolutive noise, and illumination level has not been assessed systematically. This paper presents an experimental evaluation of the combination of Gabor wavelets with graph matching. The investigations assess robustness against additive impulse noise, convolutive periodic noise, and ambient light level. The results show that the above pose estimation approach is sensitive to the density of impulse noise, and is mildly sensitive to illumination level. The paper also reports that pose estimation is sensitive to the amplitude of periodic noise but insensitive to its orientation.