Questions about QFL #36
Description
Hi, thanks a lot for your great work and I am sorry to bother you but there is a question that nags me a lot.
As mentioned in your paper, the supervision of the localization quality estimation is currently assigned for positive samples only which is unreliable as negatives may get chances to have uncontrollably higher quality predictions.
Merging quality prediction into classification score is fantastic. But here is the thing: Have your tried to keep the quality branch and train it with the inclusion of negative samples with proposed QFL?
I've tried this in FCOS and it turned out that centerness predictions for positive sample in minival were pretty bad. And I've taken those super parameters like initial bias prior for centerness branch into consideration but results were always bad.
Hope to hear your insights.
Thanks again.
@implus