Automatic-Image-Colorization
Automatic-Image-Colorization copied to clipboard
lukemelas
Can't get the right output images
Dear Luke, thanks a lot for your work. It's really inspiring!
However, when I tried to train the color model by myself using the Places365 dataset, I either get color images that are nearly gray(almost same as input gray images), or all blue, green or red(all pixels have the same blue color). Also, it reports a warning that "Color data out of range:Z<0 in 8905 pixels".. I think it implies some serious error.
Can you give me some tips to train a good colorization model? Thanks a lot!
Hello,
Sorry for the delayed response. The error is occurring during the LAB-RGB colorspace conversion, because the output values of the model are not constrained to have values that make sense on the LAB color scale. This warning is normal at the beginning of training, and will slowly go away as the model learns to not produce ridiculous values. Alternatively, you could clip the values so that you don't get that warning.
Hope that makes sense, and thanks for checking out the repo!
On Mon, Jun 4, 2018, 7:47 AM BJTUJia [email protected] wrote:
Dear Luke, thanks a lot for your work. It's really inspiring!
However, when I tried to train the color model by myself using the Places365 dataset, I either get color images that are nearly gray(almost same as input gray images), or all blue, green or red(all pixels have the same blue color). Also, it reports a warning that "Color data out of range:Z<0 in 8905 pixels".. I think it implies some serious error.
Can you give me some tips to train a good colorization model? Thanks a lot!
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/lukemelas/Automatic-Image-Colorization/issues/1, or mute the thread https://github.com/notifications/unsubscribe-auth/AMsOMBUVyraHtHzf8nJYPhlFd9F5hBLLks5t5R5tgaJpZM4UY8Kc .
I faced the same issue. The final coloured images either pick up shades of blue or green, nothing else. Can you please provide a suggestion for this?
I also faced the same issue. Have someone ever solved the problem?
