search

WangWenhao0716 / Adapted-Center-and-Scale-Prediction

Pytorch implementation of "Adapted Center and Scale Prediction: More stable and More Accurate"
https://arxiv.org/abs/2002.09053
64 stars 12 forks source link

The value of weight in Figure.3? #8

Closed Ning5195 closed 4 years ago

Ning5195 commented 4 years ago

When output the parameters of ACSP(Vanilla L1).pth. For example, the value of mean_weight and var_weight are shown as follows:

layer3.17.sn3.mean_weight : [ 1.5857, -0.8275, -0.7425] layer3.17.sn3.var_weight : [-1.3445, 0.1184, 1.1824]

Is this the weight in Figure.3? If so, how did you calculate the proportion? Why the sum of mean_weight or var_weight is 0 not 1? I think I misunderstood, but I don't know the right answer. Could you please give me an explanation? Thank you for your help!

WangWenhao0716 commented 4 years ago

It may be: aIN + bLN + cBN, then: a/(a+b+c) and so on...

Now, I do not think this explanation is meaningful. Please do not study this paper.

Ning5195 commented 4 years ago

Thank you for your answer. Does it mean, for example: 1.5857/(1.5857-0.8275-0.7425) or 1.5857/(1.5857+0.8275+0.7425)

By the way, the paper is really concise and helpful.

WangWenhao0716 commented 4 years ago

Maybe it is not right. Please refer to https://arxiv.org/pdf/1907.10473.pdf, Fig. 4.

Ning5195 commented 4 years ago

Both 1.5857/(1.5857-0.8275-0.7425) and 1.5857/(1.5857+0.8275+0.7425) are wrong? All right. Thank you for your answer! Have a good day!

WangWenhao0716 commented 4 years ago

OK! I think both of them are wrong.