Use of Average Pool to Conv5_4 layers before APN?

[jeong-tae]

Use of Average Pool to Conv5_4 layers before APN?

[jtlee90]

Thank you for reaching me!

Well... I made this code long time ago... so I am not sure that why I am pooling over the feature map of conv5_4.

I remember that I followed original implementation of RACNN, which rewritten in... maybe caffe?
Yes, yes, I saw some code, *.proto from original implementation and implement into pytorch code!

Also, in the paper section 3.4, they mention conv5_4 as highest response map and last conv layer in vgg to consider squares. Maybe that's the exact reason.

[pshroff04]

Thanks for the response!

Yes, if layers are corresponding to their caffe implementation, that does make senses.

Also, I had a thought regarding the pre-training APN network. Should we keep the parameters of VGG fixed and train the parameters corresponding to APN layers only? since pre-training aims only to learn function to transform feature maps (conv5_4) to [tx,ty,tl].

Please let me knows your view on this? @jtlee90

[jeong-tae]

I guess... yes. That's the alternative training for two network.
and I think if APN already propose good enough attention, you don't need to train APN anymore.

but maybe there will be small rooms to improve attention more.