Yang

I think so and the gradient of Wi can be computed by: dLoss/dwi=gi(dxi-1/dwi)

Hi, have you achieved better performance right now?

so... why the dim0 of key_rel_w/h should be 2*self.shape-1 instead of others?

having a question and need some help... why the size should be 2*self.shape - 1?

你好我想问下看了论文后我的理解这个target IOU和predicted IOU是都在变的吗？targetIOU根据4个回归的值计算出框然后与gtbox计算得到的，不知道这个理解对不对

so...does the image must be normalized before these computation？

Thank s for your great job! It's also hard for me to reproduce the results . Can you share your dataset with me? My email is [email protected]

如果一个proposal的label assignment按照这种分配方式的话，会不会存在一种潜在的问题呢？就是proposal的第一部分的特征和输出结果其实与GT2相关度比较高，但是分给他的label是GT1，从而影响了模型训练