0
$\begingroup$

I have a dataset that has 32x32x3 images. However, I want to use models that were developed for 224x224x3 images, e.g. resnet. A common theme I see is that people resize 32x32x3 to 224x224x3, but this means a CNN is scanning large segments of identical and redundantly colored space. Conceptually, this is not ideal.

One idea is to put a few layers at the beginning of the network, allowing the network to learn a mapping from 32x32x3 to 224x224x3. I have read a little bit about deconvolutions and think that might be the answer.

What is a reasonable set of deconvolution layers to go from 32x32x3 to 224x224x3?

$\endgroup$
0
$\begingroup$

Gaussian upsampling is one way to obtain soft, larger images http://www.cse.psu.edu/~rtc12/CSE486/lecture10.pdf . It basically consists of a simple resize followed by applying a softening convolution that makes the image smoother. Also, you don't necessarily need to go all the way up to 224. In many contexts, placing a small image in the middle of a black frame to achieve some pre-trained network's dimensions works well. You might try different balances between upsampling and extension with black pixels.

Lastly, what about another network? Pre-trained networks of many sizes and kinds are available online, for example here https://modelzoo.co/ . Hope that helps!

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ The black frame seems to give the same result as upscaling. I'll look into what a softening convolution might look like in this context. I think creating a parametric mapping from 32x32x3 to 224x224x3 might still be necessary. $\endgroup$ – Joseph Konan Feb 28 at 18:02

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.