In the file: bbox_transform.py, you only set bbox_inside_weights=1 to the positive anchors.
Why used the bbox_inside_weights and bbox_outside_weights? The original paper did not mention them.
And, because of bbox_inside_weights, it seems that there will be something wrong when calculate the _smooth_l1_loss in RPN loss.
In the file: bbox_transform.py, you only set bbox_inside_weights=1 to the positive anchors. Why used the bbox_inside_weights and bbox_outside_weights? The original paper did not mention them. And, because of bbox_inside_weights, it seems that there will be something wrong when calculate the _smooth_l1_loss in RPN loss.