loss function

[ekorman]

Hi, I'm confused with the (supervised) RPN loss function used in your paper (equation 3). It looks like there is no penalty for an anchor box being positive when there isn't an object. Take e.g. the extreme case of an image with no groundtruth objects in it, then there is no penalty for an anchor having high p_i.

Contrast this with the loss in the Faster RCNN paper which does impose a penalty in this case.