Lucas Beyer, Vision Transformer researcher, details his 2013 experiments using von-Mises loss to predict continuous angles from discrete labels

First, I found a dataset with head-angle annotations, "TownCentre". This was sometime around 2013, so deep learning just started entering vision. I was the first to train a net for this.

First, the simple baselines: linear regression (on pixels in the bounding-box), then add some conv+pool layers, then learn that modulo is useless in a loss function.