bbox_inside_weights & bbox_outside_weights for RPN loss

[mathild7]

Hi,
I realize this repo is no longer actively maintained, but I would like to understand your reasoning behind these variables in the RPN loss calculation.

bbox_inside_weights
Purpose:
Used as a mask to select the best anchors generated.

Issue/Question:
http://www.telesens.co/2018/03/11/object-detection-and-classification-using-r-cnns/
states that negative examples(background) should not be used in the loss calculation, but you do. Is this a mistake or did you find that this yielded better results?

bbox_outside_weights
Purpose:
Added mechanism to bias the loss towards positive or negative examples.

Issue/Question:
This mechanism also divides out the regression loss values, but you do this a second time with:
loss_box = loss_box.mean()
So this doubles the division by total number of samples. Is this a mistake or intended as a way to stabilize the loss function? You dont appear to do this with the second stage loss (bbox_outside_weights is a duplicate of the bbox_inside_weights tensor).

Thanks!