Human head pose estimation in images has applications in many fields such as\nhuman-computer interaction or video surveillance tasks. In this work, we\naddress this problem, defined here as the estimation of both vertical\n(tilt/pitch) and horizontal (pan/yaw) angles, through the use of a single\nConvolutional Neural Network (ConvNet) model, trying to balance precision and\ninference speed in order to maximize its usability in real-world applications.\nOur model is trained over the combination of two datasets: 'Pointing'04'\n(aiming at covering a wide range of poses) and 'Annotated Facial Landmarks in\nthe Wild' (in order to improve robustness of our model for its use on\nreal-world images). Three different partitions of the combined dataset are\ndefined and used for training, validation and testing purposes. As a result of\nthis work, we have obtained a trained ConvNet model, coined RealHePoNet, that\ngiven a low-resolution grayscale input image, and without the need of using\nfacial landmarks, is able to estimate with low error both tilt and pan angles\n(~4.4{\\deg} average error on the test partition). Also, given its low inference\ntime (~6 ms per head), we consider our model usable even when paired with\nmedium-spec hardware (i.e. GTX 1060 GPU). * Code available at:\nhttps://github.com/rafabs97/headpose_final * Demo video at:\nhttps://www.youtube.com/watch?v=2UeuXh5DjAE\n