calculating localization loss

[daodaofr]

trafficstars

https://github.com/chenyuntc/simple-faster-rcnn-pytorch/blob/6a538bdd113bab649f388aa48193c1b1ccc48ee6/trainer.py#L257

I find a small issue. why in this line gt_label can be equal to zero? loc_loss /= ((gt_label >= 0).sum().float()) # ignore gt_label==-1 for rpn_loss

But in Line 254, gt_lable can only be greater than 0: in_weight[(gt_label > 0).view(-1, 1).expand_as(in_weight).cuda()] = 1

[pangkun248]

I am also confused about the code in this place, and if this is modified, mAP will not be as good as before. Have you figured out why it is written here?

[CloudShen-107]

Because gt_label has elements as -1.