0
$\begingroup$

I need some help calculating the feature vector size for this network I'm working on from a paper.

It says the input image is 96x96x1.

The first convolution layer is a filter of 5x5 with a stride of 2, depth of 64. Followed by ReLU, followed by a maxpool of a 2x2 filter with a stride of 2.

Next is another convolution layer which has a filter of 5x5 with a stride of 2, depth of 128. Followed by ReLU, followed by a maxpool of a 2x2 filter with a stride of 2.

The last convolution layer has a filter of 5x5 with a stride of 2, depth of 256. Followed by ReLU, followed by quadrant pooling of a 12x12 filter with a stride of 12.

At the end, the paper's authors mentioned that they obtain a feature vector of 4032 which is their first FC layer.

Does this make sense ? Because when I try to calculate it, the last layer before the quadrant pooling will make the final dimension of the feature map 1x1x256. Which shouldn't be the case.

Any chance they're wrong with the reported parameters?

If I tried to use a stride of 1 and padding of 2 for the convolutions then it makes more sense. However I still don't get the final feature vector of size 4032 that's reported by the authors.

$\endgroup$

0

You must log in to answer this question.

Browse other questions tagged .