Abstract
Estimating the actual head orientation from 2D images, with regard to itsthree degrees of freedom, is a well known problem that is highly significantfor a large number of applications involving head pose knowledge. Consequently,this topic has been tackled by a plethora of methods and algorithms the mostpart of which exploits neural networks. Machine learning methods, indeed,achieve accurate head rotation values yet require an adequate training stageand, to that aim, a relevant number of positive and negative examples. In thispaper we take a different approach to this topic by using fractal coding theoryand particularly Partitioned Iterated Function Systems to extract the fractalcode from the input head image and to compare this representation to thefractal code of a reference model through Hamming distance. According toexperiments conducted on both the BIWI and the AFLW2000 databases, the proposedPIFS based head pose estimation method provides accurate yaw/pitch/roll angularvalues, with a performance approaching that of state of the art ofmachine-learning based algorithms and exceeding most of non-training basedapproaches.