Input and output sizes of images？

[ArtScanner]

Why do I find that the output size will be smaller in the training results？
How can I modify the code to customize the output size or the same size as the input image...

[Phillip Isola]

fineSize specifies the output image size, e.g., DATA_ROOT=./datasets/facades name=facades_generation which_direction=BtoA fineSize=256 th train.lua

[MoeinSorkhei]

Hi @phillipi,

I have a question about the Cityscapes output images in your paper. You mentioned that you evaluate the model on 256x256 images, while the visual examples that you show in your paper are rectangular rather than square.

Do you manually resize your 256x256 images into rectangular images in your paper or do you train another separate model on rectangular images?

Thanks in advance.

[Phillip Isola]

For visualization we resized to rectangular images, but the model always output 256x256.

[MoeinSorkhei]

Hi @phillipi ,
Thanks for the response.
I wanted to ask the same question about the photo->label direction.
Do you also resize the images to 256x256 for training, and then back to the original size1024x2048 for evaluation of the generated segmentations? (I could not find this detail in your paper)

Thanks in advance.

[Phillip Isola]

That's right, it is trained at 256x256. For evaluation, more information is here: 1, 2

My advice is one should almost never compare numbers between papers, only within papers. So don't worry if your eval protocol isn't exactly the same as what we did, as long as yours is reasonable. If you have a new method you are trying, then run your evaluation code on both your new method and on the baseline to directly compare the results.

[MoeinSorkhei]

Great! Thanks for your feedback!