June 2020
tl;dr: Monocular estimation of body orientation.
Overall impression
The paper annotates body orientation in COCO dataset.
TUM dataset has 8 bins, and got later extended to continuous labels by averaging annotations from 5 labelers.
Key ideas
 72 bins x 5 deg interval for annotation. This is still within the human cognition limit.
 The circular gaussian loss is used to blur the onehot classes and then regresses a heatmap. Interestingly the paper uses L2 loss directly instead of a BCE loss.

$p \propto e^{\frac{1}{2\sigma^2}(\min( 
igt 
, 72 
igt 
))}$ 
 The loss is approximation von Mises distribution (see mono3D++ which actually used this idea to regress angles in 360 bins).
 The paper also formalized the definition of body orientation by TxS, perpendicular to both torso direction and shoulder direction. This is needed to incorporate body orientation as a weak supervision to 3d body estimation.
Technical details
 Labeling tool design. It is a brilliant idea to show case example images with body orientation. (This could be done for car orientation estimation as well).
Notes
 The version with appendix can be found on amazon.
 Classification over regression: Sometimes it is useful to convert regression to a multiclass classification with ordered bins. Instead of directly predicting a onehot label, it blurs the one hot label by allowing leakage into its neighboring bins. This is similar to the idea of label smoothing to stabilize training. I feel that this is actually a particularly useful technique for multibin classification, when the bin numbers are very large, and when bins are ordered.
 Designing Deep Convolutional Neural Networks for Continuous Object Orientation Estimation proposed three different methods to regress a continuous angle.