loss function

[ascend-research]

loss function

[lucaskyle]

i think u may check their previous works
https://github.com/haofanwang/accurate-head-pose

in the paper they mentioned:

A new wrapped-loss stabilizes the network at large yaws.

Yaw prediction is divided into 120 3◦ bins covering the full range of yaws (−180◦,180◦].
Pitch and roll predictions are each made from 66 3◦ bins covering the range [−99◦,99◦], although
only the bins from [−90,90] are ultimately used.

i think they just quite the same.