Caffe: Why adding Normalize layer only after conv4_3

[alenaliu]

Hi @weiliu89
I'm so confusing about the function of Normalize layer,why just adding it after conv4_3 instead of every multibox_source_layer('conv4_3', 'fc7', 'conv6_2', 'conv7_2', 'conv8_2', 'pool6')?By the way,how to set the parameter of the normalize layer(like across_spatial, channel_shared, the value of scale_filler)
sincerely need your help

[weiliu89]

I only observed that conv4_3 has an issue if I do not normalize it. It is probably because its feature has very large value. I haven't tested normalize other layers. You could refer to some of the details form my ParseNet paper.

[zimenglan-sysu-512]

@weiliu89 why do you init the parameter scale_filler to be constant and the init value (20) so large? if i want use the NormalizeLayer, how do i set this parameters according to my task?
btw, why it need to be learnable?
thanks.

[weiliu89]

The L2 trick is only needed when your base network is not trained with BN. 20 is some empirical value that seems to work well. You can check the ParseNet paper for more details.