“…Previous research in this area has, for the most part, viewed the problem as a classification problem wherein the viewing sphere is (artificially) quantized into non-overlapping subintervals, and head pose is represented by a set of discrete pose labels-rather than a continuum of pose angles. This approach appears to be adequate for coarse pose estimation (with some reservations) [9,10,11,12,13,14,15], and other classification problems such as facial expression and face recognition [16,17,18,19,20]. It is, however, fundamentally flawed when used for fine-grain pose estimation for two main reasons: (i) pose estimation discontinuities occur at class boundaries due to the arbitrary nature of the pose classes, (ii) the numerical properties (scale, well-ordering) of the underlying pose angles are lost; for example, the difference between pose label 1 and pose label 2 is viewed no differently than between pose labels 1 and 5.…”