2D Image head pose estimation via latent space regression under occlusion settings

Head orientation is a challenging Computer Vision problem that has beenextensively researched having a wide variety of applications. However, currentstate-of-the-art systems still underperform in the presence of occlusions andare unreliable for many task applications in such scenarios. This work proposesa novel deep learning approach for the problem of head pose estimation underocclusions. The strategy is based on latent space regression as a fundamentalkey to better structure the problem for occluded scenarios. Our model surpassesseveral state-of-the-art methodologies for occluded HPE, and achieves similaraccuracy for non-occluded scenarios. We demonstrate the usefulness of theproposed approach with: (i) two synthetically occluded versions of the BIWI andAFLW2000 datasets, (ii) real-life occlusions of the Pandora dataset, and (iii)a real-life application to human-robot interaction scenarios where faceocclusions often occur. Specifically, the autonomous feeding from a roboticarm.