what caused r_fast.png in GuidedFilter_TF/test and GuidedFilter_PyTorch/test lookes so different? about deepguidedfilter HOT 8 CLOSED

Have you tried to reproduce the two images yourself? I think the image in Github might not be consistent with the code.

from deepguidedfilter.

Yep, I did. I have run the 'test_fast_guided_filter.py', but I can't get consistent result still. Here is the result of tensorflow

and this is result of pytorch

from deepguidedfilter.

I found the issue. The guided image in pytorch is the input image while it's the output image in TF.

from deepguidedfilter.

OK, they look the same now, thank you!

from deepguidedfilter.

I also ran into this issue, but i fail to see why the fast guided filter outputs look similar,
while the guided filter outputs look different. Could you elaborate your last comment?
Or if there is a code change, could you commit it? Thanks.

from deepguidedfilter.

The bug is in the test scripts:
In TF, the inputs are [lr_rgb, lr_gt, gt]];
In PT, the inputs are [lr_rgb, lr_gt, rgb]].
So, the test results are different.

from deepguidedfilter.

I see, thanks for the explanation. Do you mind merging the fix?
I guess many people run into this issue.

Also, for the pytorch guided filter i get strange output.
Maybe it is different pytorch version, etc., but i tried to fix the warnings (upsample vs interpolate, etc.),
but i could not find the reason. So the r.jpg looks like this:

Any advice for this issue?

from deepguidedfilter.

1. PR is welcome.
2. It's the correct output image. To obtain a high-quality image, please tune the hyper-params in GuidedFilter, i.e. radius and eps.

from deepguidedfilter.