Xiang Li

按理说 AP 是通过score 排序，越多正确的排在前面AP越高，所以不太应该会导致误报率变高啊，有点奇怪。。。是不是需要调整一下卡的阈值呢？

Hi, we do not observe the gradient problem when training upernet using the pre-trained model. The message is correct since segmentation model indeed doesnot need fc and head layers of...

> Thank you for your great work! I have a problem when I tried to run your model on my single computer. The quality focal loss is so big up...

config的 GFocalHead 里面 loss_cls=dict( type='QualityFocalLoss', use_sigmoid=False, beta=2.0, loss_weight=1.0), 注意use_sigmoid=False 很多对应的代码有一些变化，请注意一下~

hi： 建议你可以参考ttfnet，它已经修改为上下左右回归，而且收敛比centernet快很多； 旋转的话得看你怎么设计成一种分布式表示了，我目前也没接触过旋转这类，可能还要再细想一下~ ------------------ 原始邮件 ------------------ 发件人: "implus/GFocalV2"

Hi Each object will generate its own distributions of four sides. Please refer to more details in GFocalV1, thank you~ Best, Xiang

不正常，训练的时候没有遇到过。请问两者分别是什么训练setting？

config的 GFocalHead 里面 loss_cls=dict( type='QualityFocalLoss', use_sigmoid=False, beta=2.0, loss_weight=1.0), 注意use_sigmoid=False 很多对应的代码有一些变化，请注意一下~

You can check https://github.com/implus/GFocalV2/blob/master/mmdet/version.py for detailed version (2.6.0) and find correct INSTALL.md and GETTING_STARTED.md from mmdet official repo. We'll update them here later.

> Hi! I've been following gfl for some time, and it's a great work. > Actually, the uncertainty of bbox's distribution could produce some bad results sometime in gfl1. It...