search

uber-research / UPSNet

UPSNet: A Unified Panoptic Segmentation Network
Other
648 stars 120 forks source link

Why your bounding box regression loss is so small? #44

Closed JoyHuYY1412 closed 5 years ago

JoyHuYY1412 commented 5 years ago

I found that the bounding box regression loss at first of your code is really small, like of ~0.1 scale. Could you give me some explanation?

YuwenXiong commented 5 years ago

bbox_loss is only for fg rois, and at the beginning when RPN is not well trained there are not many fg rois.

JoyHuYY1412 commented 5 years ago

bbox_loss is only for fg rois, and at the beginning when RPN is not well trained there are not many fg rois.

So do you mean the loss is calculated among the fg rois but averaged between all the fg and bg rois?

YuwenXiong commented 5 years ago

@JoyHuYY1412 yes, see here: https://github.com/uber-research/UPSNet/blob/master/upsnet/models/rcnn.py#L187