On the power of data augmentation for head pose estimation

Deep learning has been impressively successful in the last decade inpredicting human head poses from monocular images. However, for in-the-wildinputs the research community relies predominantly on a single training set,300W-LP, of semisynthetic nature without many alternatives. This paper focuseson gradual extension and improvement of the data to explore the performanceachievable with augmentation and synthesis strategies further. Modeling-wise anovel multitask head/loss design which includes uncertainty estimation isproposed. Overall, the thus obtained models are small, efficient, suitable forfull 6 DoF pose estimation, and exhibit very competitive accuracy.