Mmdetection: Why "focal loss" use only 'num_total_pos' for 'avg_factor?

[kye9216789]

I do not have understand about anchor_head.py because of :

    num_total_samples = (num_total_pos if self.use_focal_loss else
                         num_total_pos + num_total_neg)

and I think that focal_loss also should use every sample for loss, and devide by number of every sample : num_total_pos + num_total_neg.
Please make me understand why focal loss use only num_total_pos.

[hellock]

When using focal loss, no bbox sampling is performed and the number of neg bboxes is very large. We follow the official implementation to use pos bboxes only .

[kye9216789]

Does it related in part of paper,
"The total focal loss of an image is computed as the sum of the focal loss over all ~100k anchors, normalized by the number of anchors assigned to a ground-truth box."
??
"Anchors assigned to a ground-truth box" means positive sample, is it right?

[hellock]

You got it.

[kye9216789]

Thanks for your graceful help!