converting classification network to fully convolutional with arbitrary sized image

I have trained a classification network that runs on 32x32x3 input (classify letters). How can I convert it to fully convolutional network that will run on arbitrary sized image. I don't want to resize my input to 32x32x3, but run the detection over the whole image classifying every neighborhood of 32x32x3 sequentially.

Thank you, in advance.

Answer 1

[Prateek Rai]

Hi,

One possible workaround could be that you can first use the whole image and find the potential patches of size 32x32x3 where there is a high chance of having letters.

(Note: If you want to use the entire image then you can start from the top left and take a window of size 32x32x3 and then start moving that window sidewards to cover the whole image eventually.)

Once you start having the small images of size 32x32x3, then you can use those to feed to your Convolutional Network and get the desired output.