Tensor size mismatch when testing on own png about pynet-pytorch HOT 5 CLOSED

@Garfield-kh,

The size of the input image should be multiple of 32 due to the used down- and up-scaling layers. Changing the size from 2010 x 3018 to 1984 x 3008 should fix the problem.

from pynet-pytorch.

Thank you!
Yes, it works.
May I ask one more question?
When the brightness of Raw image is very high, then the resulit RGB image will be almost in white. It this because the training data is in a low exposure setting?
I failed to download the data for training thus I cannot check it myself.
The performance is quite good on low exposure RAW image.

Thank you again.

from pynet-pytorch.

Hi @Garfield-kh,

As far as I can see, you collected RAW images with a phone model different from the Huawei P20 used for training this network. Therefore, your device has another camera sensor model and optical system. While the PyNET is still able to generalize even to such hard cases, various small artifacts can be expected when applying it to raw data from other devices.

from pynet-pytorch.

Thank you for the answering!
Yes, it's different  device.
What I want to try is to test on different exposure RAW using this model.

• When I try increase the exposure value of RAW, (scale up the RAW value 50% and clip on the testing data 'test/huawei_full_resolution/1167.png' ). It seems like the output will get some small artifacts like blue region.
• In my own different exposure RAW data, it works good for proper exposure setting, and some region of output image go blank in high exposure setting.

As the paper state 'The photos were captured in automatic mode' in the section 'Zurich RAW to RGB dataset', Does this means the pretrained model can only handle suitable exposure setting since the training only include proper exposure RAW image?

from pynet-pytorch.

I find similar issue which in SID (See in Dark), the author suggests to try and re-scale the raw with a proper ratio such that the model is suitable for it.
Thanks~

from pynet-pytorch.