Questions about choosing loss scale parameters

[minooisbusy]

Hello, and thank you for the great works.

I have some questions about the loss function.

As written in the paper, the loss for the training network uses Huber loss (pseudo-Hubber loss to be precise), and the robust function of iterative solutions uses alpha = 0.

I have a questions here about setting the scale parameters.

1. The scale parameter c is 0.1 in the previous loss and 2 in the later loss. Can you explain why you chose each scale parameter?
2. In the function reprojection_error(T_r2q), variable err is passed to scaled_barron(), dividing the returned value by 4. Is there a special reason for this value to be divided by 4?