Question about the design of QGN

[exiawsh]

Hello~ Thanks for your great work! I have some question about the label assignment in QGN. It seems that you follow the design of "End-to-End Object Detection with Fully Convolutional Network". Have you tried the label assignment proposed in OneNet (What Makes for End-to-End Object Detection?), What about the performance?

[mulinmeng]

Hello~ Thanks for your great work! I have some question about the label assignment in QGN. It seems that you follow the design of "End-to-End Object Detection with Fully Convolutional Network". Have you tried the label assignment proposed in OneNet (What Makes for End-to-End Object Detection?), What about the performance?

Thanks for your ask, we temporarily define two label-assignment methods as "sum" and "multi". The two methods are generally similar, ”multi“ has 0.1% AP gain than "sum" at some settings.