image size problem

[zmh0510]

hello, it is a wonderful project! But i am wondering that i must input a fixed image to the project? say 6363? How can i input a 600480 image to the project? thanks a lot !

[yuke93]

@zmh0510 Thanks for your comment! Yes, currently it only supports 63x63 input. We have tried a simple extension to support arbitrary size input. Specifically, we crop a large image into 63x63 patches and let each patch vote for a tool, then conducts the tool with most votes on the whole image. I plan to release this feature after updating training code.

[zmh0510]

OK！looking forward to it, especially the training code!

[hernan3009]

@yuke93 It looks very interesting. Before see this issue I think in the following simple idea:

1 - Generate many patches sharing (in some way) common pixels.
I thought in something like: If a patch cover a 63*63 pixel square from (x_0,y_0) to (x_0+62,y_0+62), the next horizontal patch going from (x_0+31,y_0+31) to (x_0+31+62,y_0+31+62) ... And something analogous for the y coordinate.
2 - Then reconstruct the image using some kind of interpolation. For example, weighting the patch contribution depending on the distance to the patch center.

In your experience, do you think that this could work? If it is plausible, how can I choose "the best" output image for the patch? I see in the output something like:

mano_in_11_13_13.png mano_in_11_13.png mano_in_11.png
8_in_5_11_13.png 8_in_5_11.png 8_in_5.png

but I am unsure about the naming convention.

Regards

[yuke93]

Hi @hernan3009 ,

I have tried some similar interpolation approaches as you said but the results are sometimes not pleasant because there is not any constraint about the selected toolchains between adjacent patches. When the selected toolchains are different, the interpolated image might look strange, and the PSNR is not as good as the voting method as I described above.

The number after '_in_' stands for the indices of sequentially selected tools. Please see the table below:

Index	Distortion Type	Distortion Level Interval
1	Gaussian blur	[0, 1.25]
2	Gaussian blur	[1.25, 2.5]
3	Gaussian blur	[2.5, 3.75]
4	Gaussian blur	[3.75, 5]
5	Gaussian noise	[0, 12.5]
6	Gaussian noise	[12.5, 25]
7	Gaussian noise	[25, 37.5]
8	Gaussian noise	[37.5, 50]
9	JPEG compression	[60, 100]
10	JPEG compression	[35, 60]
11	JPEG compression	[20, 35]
12	JPEG compression	[10, 20]
13	None	N/A

[lvzhengyi0204]

Hello, it is a exciting work, I want to know whether the test stage have the ability of arbitrary size input.
In the code, I don't see the related features. I am looking forward to your reply.

[yuke93]

Hi @lvzhengyi0204,

I haven't added this feature yet. I will implement the heuristic method, as discussed above, to support arbitrary size input, by the end of this week.

[yuke93]

Dear all,

Thanks for your advice and comments. Now we support arbitrary input size while inference. Put your images in data/test/mine/ and then run

python main.py --dataset mine

[hernan3009]

Thank you! Works here under Python3.

[lvzhengyi0204]

Thank you very much！ It helps a lot for my research work.

[yuke93]

I am very glad that you like it 😃

[s-sanyal]

I really liked this project. I recently faced an issue while using this project in Google Colab. absl.flags._exceptions.DuplicateFlagError: The flag 'log_dir' is defined twice. First from absl.logging, Second from /content/RL-Restore/main.py. Description from first occurrence: directory to write logfiles into
Could you please explain how to get rid of this issue. Thanks in advance :) @yuke93

[JialinKang]

I really liked this project. I recently faced an issue while using this project in Google Colab. absl.flags._exceptions.DuplicateFlagError: The flag 'log_dir' is defined twice. First from absl.logging, Second from /content/RL-Restore/main.py. Description from first occurrence: directory to write logfiles into
Could you please explain how to get rid of this issue. Thanks in advance :) @yuke93

I also have the same question, hope to get a reply! Thanks a lot!