About the location of BN in the head of SSD VGG-16

[LaVieEnRoseSMZ]

Hi~I am reproducing your code using pytorch. And I have gained 76.3% with BN in the Backbone. It does improve accuracy. However, I am confused about the BN in the head.

Head is made up of 3 parts which extras layers to do the feature downsampling and loc layers to predict location and conf layers to classification. Should I add BN to all the three parts or just extra layers?

I am trying to find prototxt of the actual network, but I haven't found it. Thanks in advance for your patience.