Caffe: Why normalization performed only for conv4_3?

[yeongrok]

There are several source featurs that are used for predicting confidence or localization.

Normalization is only performed just for conv4_3. Is there anyone tell me the reason?

It would be very helpful for me. Thank you in advance.

[weiliu89]

That was discovered in my other paper (ParseNet) that conv4_3 has different scale from other layers. That is why I add L2 normalization for conv4_3 only.

[yeongrok]

Oh I see. I should have read that paper. Thank you very much for your kind explanation

[ujsyehao]

HI, @weiliu89 I have two problems:

1. conv4_3 layer's bottom layer is conv4_2, so why conv4_3 has a different scale from other layers? I am confused about it.
2. I plot weight plot of conv4_3, conv4_3_norm, conv4_3_norm_mbox_loc.
   
   
   
   conv4_3 layer weight range from -0.1 to 0.25, then conv4_3_norm layer do normalization, and its weights range from 13.20 to 13.30, so what is meaning?

[ujsyehao]

I plot 6 layers (SSD method choose six layers):

The conv4_3 figure does look different from other figures, it has a different scale.

[kangyiS]

Dear ujsyehao, how do you plot this 6 layers? And what is the meaning in each plot? The y axis is weight value, but I don't know the meaning of x axis. Thank you.

[espectre]

@ujsyehao I have the same confusion as @kangyiS ,could you have some time to help us,thaks!