NosePose: a competitive, landmark-free methodology for head pose estimation in the wild

We perform head pose estimation solely based on the nose region as input, extracted from 2D images in unconstrained environments. Such information is useful for many face analysis applications, such as recognition, reconstruction, alignment, tracking and expression recognition. Using the nose region has advantages over using the whole face; not only it is less likely to be occluded by acesssories, it is also visible and proved to be highly discriminant in all poses from profile to frontal. To this end, we propose and compare two different approaches, based on Support Vector Machines (SVM-NosePose) and on Convolutional Neural Networks (CNN-NosePose) such that no landmarks are needed to perform pose estimation, favoring success in extreme pose and environment where landmark detection is non-trivial. Our NosePose methodology was applied to four publicly available uncontrolled image datasets (McGillFaces, AFW, PaSC and IJBA). Results show that both SVM-NosePose and CNN-NosePose approaches are competitive, through thoughtful and comprehensive experiments, when compared against state-of-the-art works on head pose estimation. Keywords-Head pose estimation; Nose pose estimation; Face image analysis; Support vector machines; Convolutional neural network