I'm trying to perform a segmentation task on images of multiple sizes using fully convolutional neural networks.

Currently, I'm using EfficientNet as a feature extractor, and adding a deconvolution/backwards convolution/transposed convolution layer as described in the original Fully Convolutional Networks for Semantic Segmentation paper.

But this transposed convolution layer doesn't return a filter of a size equivalent to the original image for images of varying sizes.

For example, let's assume the original image is $100 \times 100$, and the last layer contains filters of size $50 \times 50$. To get a filter of the same size as the original, you would need a transposed convolution layer of size $51 \times 51$.

Now, assume you passed in an image of size $200 \times 200$. The last layer would contain filters of size $100 \times 100$. That same transposed convolutional filter of size $51 \times 51$ would result in an output of size $150 \times 150$.

Is there any way to make it so that a fully convolutional network always returns an image of the same size as the original?

  • $\begingroup$ Have you solved this issue? If yes, how? You may want to write a formal answer below, for reference. $\endgroup$ – nbro Jun 14 at 11:03

I ended up using a work around.

I set up the network so that an C x C (i.e. 320 x 320) input would output a C x C mask for some constant C (in my case it was 320).

I then resized the image I wanted to pass in to C x C, and then resized the output back to the original size of the Image.

| improve this answer | |

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.