Some question about the EfficientAD PDN structure

[1094558673]

Hello, I have some questions about the EfficientAD PDN structure: in my opinion, if the input is 3256256, the output is 384×64×64; Then how to represent the 33×33×3 Input Patch in PDN? I see the code is just a simple nn.Conv2d operation.

[abc-125]

Hello, please take a look at the Figure 2 in the EfficientAD paper. 33×33×3 Input Patches become 1×1×384 Patch Descriptors, so in the output you have results for many overlapping patches. In the code it is represented by nn.Conv2d followed by nn.AvgPool2d .

[1094558673]

Hello, please take a look at the Figure 2 in the EfficientAD paper. 33×33×3 Input Patches become 1×1×384 Patch Descriptors, so in the output you have results for many overlapping patches. In the code it is represented by nn.Conv2d followed by nn.AvgPool2d .

Thank you very much for your reply. I understand what you mean. I have also seen the content of Figure 2 in the paper. But I see in the code that the input size of his image in PDN should be 256x256, how to change to 33x33, I only see nn.Conv2d and nn.AvgPool2d in the code, and do not see other operations.

[blaz-r]

Hello. The 33x33 patches are sort of implicit due to the way convolution works. You can check this page for more info on receptive field of CNNs.