search

longcw / faster_rcnn_pytorch

Faster RCNN with PyTorch
MIT License
1.71k stars 466 forks source link

why loss_box is multiplied with 10 to calculate the final loss? #54

Closed zchrissirhcz closed 2 years ago

zchrissirhcz commented 7 years ago

In both RPN subnetwork and RCNN subnetwork, the loss in your repo, is both defined as:

@property
def loss(self):
    return self.cross_entropy + self.loss_box * 10

This is different from the original paper's formula. Why x10 on bbox loss, instead of just adding cross entropy loss and bbox loss?