Questions about the backbone

[DreamerrW]

hi, i'm learning your code,and i find a quesion in your PFENet.py file.

in line 204:
if self.vgg:
query_feat_2 = F.interpolate(query_feat_2, size=(query_feat_3.size(2),query_feat_3.size(3)), mode='bilinear', align_corners=True)

I want to ask that why the Resnet's out of conv2_layer and conv3_layer don't need interpolation to make their Feature maps the same size.

thank you for your reply!

[tianzhuotao]

Because the ResNet has used dilated convs that make the outputs of the last three layer blocks have the same spatial size, i.e., 1/8 of the input image's size.

[DreamerrW]

I get it，Thank you！