The question about that the pos weight of each anchor for each instance is normalized by the sum of all pos weights within the candidate bag.

[MooresS]

Hello, I take the liberty to ask. The pos weight of each anchor for each instance is normalized by the sum of all pos weights within the candidate bag.

p_pos_weight = (torch.exp(5*p_pos) * p_pos * center_prior_weights) / (torch.exp(3*p_pos) * p_pos * center_prior_weights).sum(0, keepdim=True).clamp(min=EPS)

In this code, why is the numerator 5*p_pos but the denominator 3*p_pos?

[strongwolf]

This setting mainly aims to adjust the balance between pos loss and neg loss. You can refer to this issue.

[MooresS]

If so, is it normalized？Or it multiplies exp(2*p_pos) after normalized.

[strongwolf]

Yes, it can be seen that pos_weight multiplies an extra term after normalization.

[MooresS]

Thanks for your reply.