SeungjunNah / DeepDeblur_release

Deep Multi-scale CNN for Dynamic Scene Deblurring
671 stars 142 forks source link

diff image is zero-valued #64

Closed haow85 closed 2 years ago

haow85 commented 3 years ago

Dear code contributors:

I tested the algorithm package using the following command :

python demo.py --gpu 0 --train_dir pretrained --exp_name multi_skip --image "photo.jpg"

, and in the function compute_psnr() of utils.py, I got a math domain error due to Inf value for log10. The reason is because img1 - img2 is zero valued array, which means img1 and img2 are the same.

How do I fix the issue ?

Best Regards, Hao Wang

haow85 commented 3 years ago

I found in the test function of demo.py the input image directory is required to contain an input_b1 image and a target_s1 image. I think they are referring to blurry image and sharp image, but in my test case I do not have a target sharp image for performance evaluation.

I commented out the line compute_psnr and the print function below. The error disappeared, but when I took a look at the resulting deblurred image, I got a blank white canvas with a broken yellow line.

I think the algorithm failed in the case.

SeungjunNah commented 3 years ago

Hi @haow85,

This is a torch version repository that uses the Lua language and you are using a python command. Also, based on the input argument you used, the code you used doesn't seem to be related to the Pytorch version of this code repository.

Could you check if you are referring to this repository?