Image Matting about shapematchinggan HOT 7 CLOSED

Maybe your image contains the alpha channel, i.e., RGBA channels, while our code works on RGB channels. So there is a mismatch in the channel dimension (4 v.s. 3).
In addtion, the length and width of the image need to be a multiple of 8. For example, 255*255 should be resized to 256*256.
Or you can provide me with a link of your image and I will check it.

from shapematchinggan.

I just used Photoshop. It is easy to use and the user can control where to be used as foreground.

from shapematchinggan.

Thank you for such a quick response. I am trying to train the model on a different dataset but I have no background in photoshop. Can you please suggest steps you followed to convert the text data used in TET-GAN paper to data used in this paper? Any steps or reference link would be really helpful!

from shapematchinggan.

For a new style image, you may refer to the magic wand tool in photoshop (https://www.photoshopessentials.com/basics/selections/magic-wand-tool/)
to select its foreground. Then you fill the foreground with white and the remain region with black.
Next, use utils.text_image_preprocessing to transform black and white mask into distance-based mask (the blue and red one).
FInally, concatenate distance-based mask with your style image as the format of images in ../data/style/ and copy the result to ../data/style/.
Hope this help.

from shapematchinggan.

@williamyang1991 Thank you for this awesome project! I've followed these instructions and have obtained an image in the same format as the those in ../data/style/ (distance map on the left, original style reference on the right) and the network nearly makes it through the first epoch before throwing

"RuntimeError: The size of tensor a (256) must match the size of tensor b (255) at non-singleton dimension 1"

I'm almost certain it has to do with the way I may be formatting the image, but I'm unsure. Any ideas?

from shapematchinggan.

Changing the length & width of my image to multiples of 8 fixed the problem. Thank you!

from shapematchinggan.

That was really helpful! Thanks for the Photoshop instructions!

from shapematchinggan.