I have not seen this explicitly stated anywhere so I was curious. Say I have network trained to meet my segmentation needs using 250x250 images. After this training is complete and I wish to submit images in production for segmentation, do those production submitted images need to be 250x250 as well or can they be any reasonable size?

If they must be resized to 250x250 for segmentation, is it possible to scale up the segmentation regions to apply to a larger image? If so what is the name of that technique so I can research it a bit more.


1 Answer 1


Yes, you will need data that has the same dimensionality as your training set. Otherwise the model will either discard excess data, or will supplement missing data with a random number. This will likely lead to very poor results.

I would suggest to rescale your images to have the same pixel density as those you used for training (250x250) rather than scaling your model.

Methods for rescaling images and maintaining details are very well developed. Many exist and you can use existing tools such as Photoshop to change the pixel density of an image.

  • $\begingroup$ Yup, I can totally take the image and scale it down to the training size no problem. I'm attempting to use 3 class segmentation to build a rough image matte so while scaling it down to get the segments is no problem, I will need to scale that back up to properly matte the image. Scaling up while keeping a clean matte would be the trick. $\endgroup$ Commented Jul 13, 2017 at 14:57
  • 1
    $\begingroup$ Or I guess my other option is to train it on the image sizes I expect to feed it in production. $\endgroup$ Commented Jul 13, 2017 at 14:59
  • $\begingroup$ It is always best practice to train your model using data as similar to what you will be expecting on the field. Ideally, your data should be drawn independent and identically distributed (iid) from the real world distribution you will be using the model with. $\endgroup$
    – JahKnows
    Commented Jul 13, 2017 at 15:01

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .