Yolov5: Objectness loss

Created on 23 Oct 2020 · 2Comments · Source: ultralytics/yolov5

❔Question

In the compute_loss() method, when calculating the objectness loss, only calculate the objectnessloss of the positive sample? Did the negative sample not participate in the calculation?

Additional context

tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * iou.detach().clamp(0).type(tobj.dtype)
lobj += BCEobj(pi[..., 4], tobj) * balance[i] # obj loss

Stale question

Source

zxj11838

👍1

Most helpful comment

I think that objetcness loss using all sample, not only positive sample.
The variable pi means all samples in a prediction layer.

Building the GT label corresponding to pi.

tobj = torch.zeros_like(pi[..., 0], device=device)  # target obj
tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * iou.detach().clamp(0).type(tobj.dtype)  # iou ratio

Calculating loss

lobj += BCEobj(pi[..., 4], tobj) * balance[i]  # obj loss

WZMIAOMIAO on 23 Oct 2020

👍2

All 2 comments

I think that objetcness loss using all sample, not only positive sample.
The variable pi means all samples in a prediction layer.

Building the GT label corresponding to pi.

tobj = torch.zeros_like(pi[..., 0], device=device)  # target obj
tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * iou.detach().clamp(0).type(tobj.dtype)  # iou ratio

Calculating loss

lobj += BCEobj(pi[..., 4], tobj) * balance[i]  # obj loss

WZMIAOMIAO on 23 Oct 2020

👍2

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.