vcmod a plugin for vapoursynth

[vcmohan]

vcmod is a plugin for vapoursynth. It has several functions which essentially operate on pixel values and modify it as per some criteria. Useful for denoising.
The functions are

1.Amplitude
2.Histogram
3.Median
4.Variance
5.Gaussian Blur GBlur
6.Motion Blur MBlur.

A full description is available at vcmod (http://www.avisynth.nl/users/vcmohan/vcmod/vcmod.html).
Request comments. This plugin is compiled for windows.

[mawen1250]

http://i.imgur.com/tUOkuuh.png
(4x)

I've got these artifacts with vcmod.Median(clip, 7, [1,1,1]).
Is it expected by the algorithm?

[vcmohan]

It is not expected. Will it be possible to attach your input one frame and send me by email? I can checkup easily then.

[mawen1250]

That's OK. Already sent.

[vcmohan]

An updated version is at my page. Please check it .

[cwk]

@vcmohan, Is the neural net function included with your vcmod plugin? The web page (http://www.avisynth.nl/users/vcmohan/vcmod/vcmod.html) does not reflect its presence.

[vcmohan]

Sorry. (I have not yet ported it to vcmod. It is currently available in modplus a avisynth+ plugin. If you are particular I will port it to vapoursynth in a week.

[cwk]

Your comment on the avisynth thread about sharpening and denoising sparked my interest, though I only use Vapoursynth these days. I would like to try using a different source of the same video as a training clip for some old VHS captures.

[vcmohan]

I am almost done. But I am getting a error message "invalid plane requested" from vseditor. I am unable to see why this message. Still trying to check.

[Myrsloik]

I am almost done. But I am getting a error message "invalid plane requested" from vseditor. I am unable to see why this message. Still trying to check.

Most likely you're calling getStride() or getReadPtr() with a plane argument that's out of range. For example plane=1 for grayscale.

[vcmohan]

I have now uploaded the new vcmod having saltpepper, fan veed and neural functions additionally. Please check for any bugs or misinfo typos etc;

[ChaosKing]

broken file upload? vcmod64.7z is 0 bytes

//Edit: works now, thx.

[vcmohan]

Sorry for that. I uploaded again. Kindly check.

[cwk]

Thanks for updating vcmohan. Do you have any guidance for building on linux?

[Are_]

sed -i -e s/vapoursynth.h/VapourSynth.h/g -e s/vshelper.h/VSHelper.h/g *.cpp

gcc -O3 -msse2 -fPIC -shared -std=c++11 -Wl,-O1 -Wl,--as-needed -o libvcmod.so vcmod.cpp

[vcmohan]

Do you have any guidance for building on linux? Frankly I have no experience. I remember some one tried to transfer to GCC. He got lots of errors and warnings. While I do not care for most of the warnings VC++ (community version 13) generates, as these conversions are intended (for example int to float etc; ) probably GCC may not allow. You may pose this question to experts.

[cwk]

sed -i -e s/vapoursynth.h/VapourSynth.h/g -e s/vshelper.h/VSHelper.h/g *.cpp

gcc -O3 -msse2 -fPIC -shared -std=c++11 -Wl,-O1 -Wl,--as-needed -o libvcmod.so vcmod.cpp

Thanks Are_

Builds without issue. I'll give it a try this weekend.

[cwk]

I'm looking at the file generated during training

iter:9980
4.29846e+11 4.2984e+11 4.29845e+11 4.2984e+11 4.29845e+11
4.29839e+11 4.29845e+11 4.29839e+11 4.29845e+11 4.29838e+11
4.29844e+11 4.29838e+11 4.29844e+11 4.29838e+11 4.29844e+11
4.29838e+11 4.29845e+11 4.29838e+11 4.29844e+11 4.29837e+11

It looks like Neural generates of convoltion matrix of: xpts * ypts, but the error is calculated against 5 * 5. Is that correct?

[vcmohan]

It looks like Neural generates of convoltion matrix of: xpts * ypts, but the error is calculated against 5 * 5. Is that correct?

No. You are looking at error of covolution at iter 9980 to 9999. A convolution matrix of xpts*ypts + 1 is generated. The one that gives least error is selected and used.
It looks you are using a very large number of iterations iter. Instead it may give a better result if you try bestof of 5 to 10 and iter of 100 to 300.
The reason is that sometimes it will not be able to get off a local minimum. Also several times it may not converge to a good solution.

[cwk]

ok, thank you for the explanation. I'll try larger values of bestof.

One observation is that the training process is single-threaded. Users could potentially use larger values for both iter and bestof if all cores could be used.

[vcmohan]

ok, thank you for the explanation. I'll try larger values of bestof.

One observation is that the training process is single-threaded. Users could potentially use larger values for both iter and bestof if all cores could be used. Training occurs in the plugin init section which is guarenteed to be single threaded. As the previous convolution error is used to adjust weights which are used for next convolution, iterations can not be multi threaded. Neural Networks Learning process is therefore slow. This is the reason option to save weights and other parameters is provided. Next time for a very similar input data those weights can be used by read file option.