Implementation doesn't match description in Paper

[dominikandreas]

In the paper L2Norm is applied before the Deconvolution of the different feature stages, in the implementation it's the other way around. It may not make a big difference, but it's probably more efficient to do the normalization first (as described in the paper). Edit: actually not if the input channels for normalization >> 256. But maybe the authors can comment on this discrepancy and either update the code or the paper

[dominikandreas]

Another difference to the paper is the score threshold during inference. In the paper 0.01 is mentioned, the code uses a threshold of 0.1. @liuwei16 could you comment on this?

[Boisa]

Yes，I also noticed this problem，Have you got a good answer?