On the representation and methodology for wide and short range head pose estimation

Head pose estimation (HPE) is a problem of interest in computer vision toimprove the performance of face processing tasks in semi-frontal or profilesettings. Recent applications require the analysis of faces in the full360{\deg} rotation range. Traditional approaches to solve the semi-frontal andprofile cases are not directly amenable for the full rotation case. In thispaper we analyze the methodology for short- and wide-range HPE and discusswhich representations and metrics are adequate for each case. We show that thepopular Euler angles representation is a good choice for short-range HPE, butnot at extreme rotations. However, the Euler angles' gimbal lock problemprevents them from being used as a valid metric in any setting. We also revisitthe current cross-data set evaluation methodology and note that the lack ofalignment between the reference systems of the training and test data setsnegatively biases the results of all articles in the literature. We introduce aprocedure to quantify this misalignment and a new methodology for cross-dataset HPE that establishes new, more accurate, SOTA for the 300W-LP|Biwibenchmark. We also propose a generalization of the geodesic angular distancemetric that enables the construction of a loss that controls the contributionof each training sample to the optimization of the model. Finally, we introducea wide range HPE benchmark based on the CMU Panoptic data set.