Denoise low-light DV with fft3dfilter, pls criticize my settings

[Vetal]

Hello, I tried to denoise my video,

http://vetals.com/Misc/Sigma1_5_Sharpen1.png

It is Sony DCR-TRV25 DV camera, in-door shooting. Before denoise is on the left, after denoise is on the right

I used:

#deinterlace
LoadVirtualdubPlugin(VirtualDub_plugin_directory+"\smart.vdf", "_VD_SmartDeinterlace", 1)
_VD_SmartDeinterlace(0, 3, 0, 100, 0, 0, 0, 0, 0, 0, 1)

#denoise
loadplugin("C:\Program Files (x86)\AviSynth 2.5\plugins\FFT3DFilter.dll")
fft3dfilter(sigma=1.5, bt=5, bw=32, bh=32, ow=16, oh=16, sharpen=1.0).



I tried different sigma and sharpen values and it looks best for my unprofessinal eye. As a compromise between noise and details

1. What fft3dfilter settings would you recommend?
2. What other denoisers would you use
3. Is fft3dgpu is absolutely the same as fft3dfilter (functionally and by parameters list), so it is processed on GPU instead of CPU?

Thank you

[G_M_C]

The denoising looks good. But as an alternative you ould experiment with HDR AGC: http://forum.doom9.org/showthread.php?p=645694#post645694 and stop using the sharpening.

[Dark Shikari]

FFT3DGPU is limited to bt=4, but its not a big deal.

My favorite is MVTools-compensated FFT3DGPU:

backward_vectors = source.MVAnalyse(isb = true, truemotion=false, delta = 1, idx = 1,overlap=4)
forward_vectors = source.MVAnalyse(isb = false, truemotion=false, delta = 1, idx = 1,overlap=4)
forward_compensation = source.MVCompensate(forward_vectors, idx=1)
backward_compensation = source.MVCompensate(backward_vectors, idx=1)
interleave(forward_compensation, source, backward_compensation)
fft3dgpu(sigma=1.5, bt=3, bw=32, bh=32, ow=16, oh=16, sharpen=1.0)
selectevery(3,1)

[TheRyuu]

I'd try a different deinterlacer like yadif which may yield better quality.

Also, I like a MVTools-compensated fft3dgpu, but I'd also recommend you try dfttest (http://forum.doom9.org/showthread.php?t=132194&highlight=dfttest) as well.

Then try throwing a LimitedSharpenFaster on that, or even SeeSaw if the original quality is high enough. I'd try avoiding using the built in sharpening of fft3d, I've found that using limitedsharpenfaster normally results in a more "pleasing to the eye" quality, but that's just me.

You decide for yourself.

[G_M_C]

FFT3DGPU is limited to bt=4, but its not a big deal.

My favorite is MVTools-compensated FFT3DGPU:

backward_vectors = source.MVAnalyse(isb = true, truemotion=false, delta = 1, idx = 1,overlap=4)
forward_vectors = source.MVAnalyse(isb = false, truemotion=false, delta = 1, idx = 1,overlap=4)
forward_compensation = source.MVCompensate(forward_vectors, idx=1)
backward_compensation = source.MVCompensate(backward_vectors, idx=1)
interleave(forward_compensation, source, backward_compensation)
fft3dgpu(sigma=1.5, bt=3, bw=32, bh=32, ow=16, oh=16, sharpen=1.0)
selectevery(3,1)

Looks like you've got resources / time / patience enough :p :D

But using FFT3DGPU is a good idea, even on my C2D system the GPU-version is faster than the CPU-version.

But vetal: Could you put up a sample-clip up somewhere, to help expirimenting ?

[Vetal]

FFT3DGPU is limited to bt=4, but its not a big deal.

My favorite is MVTools-compensated FFT3DGPU:

backward_vectors = source.MVAnalyse(isb = true, truemotion=false, delta = 1, idx = 1,overlap=4)
forward_vectors = source.MVAnalyse(isb = false, truemotion=false, delta = 1, idx = 1,overlap=4)
forward_compensation = source.MVCompensate(forward_vectors, idx=1)
backward_compensation = source.MVCompensate(backward_vectors, idx=1)
interleave(forward_compensation, source, backward_compensation)
fft3dgpu(sigma=1.5, bt=3, bw=32, bh=32, ow=16, oh=16, sharpen=1.0)
selectevery(3,1)

Very nice results! Thank you! MVTools are too advanced for me to create scripts myself.

And it hard to do side by side for test with StackHorizontal(Vertical) due to different time line branches in AVS script.
Need to use 2 VDubs for that.


But vetal: Could you put up a sample-clip up somewhere, to help expirimenting ?


I have to split it first. Smallest file I have is 530Mb. I don't think anyone would like to download it :) Never tried to split DV before, I'd better shoot some indoors 5-10 sec scene soon.

Thank you!

[R3Z]

Very nice results! Thank you! MVTools are too advanced for me to create scripts myself.

And it hard to do side by side for test with StackHorizontal(Vertical) due to different time line branches in AVS script.
Need to use 2 VDubs for that.



I have to split it first. Smallest file I have is 530Mb. I don't think anyone would like to download it :) Never tried to split DV before, I'd better shoot some indoors 5-10 sec scene soon.

Thank you!

Instead of StackVertical/Horizontal try Interleave(clip1.Subtitle("Clip1"),clip2.Subtitle("clip2")) for a direct comparison.

[TheRyuu]

Instead of StackVertical/Horizontal try Interleave(clip1.Subtitle("Clip1"),clip2.Subtitle("clip2")) for a direct comparison.

Or you can use avsp and then just use the mouse wheel to go between the different tabs :p

[R3Z]

Or you can use avsp and then just use the mouse wheel to go between the different tabs :p

More than one way to skin a cat you say ? :)

[G_M_C]

[...]
I have to split it first. Smallest file I have is 530Mb. I don't think anyone would like to download it :) Never tried to split DV before, I'd better shoot some indoors 5-10 sec scene soon.

Thank you!

I regularly get much larger files off the net ;)

But have you tried HDR Agc yet ? Just give it a try, could it can help to "lighten up" the clip, and get the colors back (less washed out cause of low light conditions).

If it doesnt work, you can allways remove it from the script ;)

[2Bdecided]

I'm guessing you want a free solution, but NeatVideo is really good for typical DV camcorder noise, and can maintain noise-like features (or very small features) that other denoisers eat very easily.

I'd second HDR AGC.

It's also worth mentioning that, if you're heading back onto DVD to watch on a TV, that amount of noise is going to be almost invisible.

If you are heading back to interlaced, make sure the deinterlacer keeps the original lines, and that you use those preserved lines in the final output. No point using interpolated output when you don't have to: it can seriously soften the image.

Cheers,
David.

[Vetal]

I regularly get much larger files off the net ;)

But have you tried HDR Agc yet ? Just give it a try, could it can help to "lighten up" the clip, and get the colors back (less washed out cause of low light conditions).

If it doesnt work, you can allways remove it from the script ;)

Here it is, http://vetals.com/Misc/Sample.avi , 274Mb

HDR Agc is not actually a denoiser as I initialy thought! That's why I was reluctant to try it - got lazy after I satisfied with another denoising method.

Result is great, but only when it goes last (Deinterlace-Denoise-HDR). Even if some noise left, it removed by x264 compression, as long as it goes not much less below CRF 20. Where it preserving such details as a noise :)

In different order, if HDR AG is before the denoiser, it is increasing noise.

I was afraid first that it would make black areas gray and bright as clipping-bright. But actual effect on sunny shooting effect is invisible. Really nice, thanks! Will include as must-have in my DV processing.

[2Bdecided]

You won't like this advice, but...

1. Buy a new camcorder - preferably an HD one
2. Use more light

One or the other - preferably both.

If it's your child: I know it's an expensive time already, but they're only this young once!

Your camcorder has a typical annoying habit with consumer DV of hunting for the focus in low light, causing it to go really blurry every few seconds. It's really annoying, and there's no fix for it in AVIsynth or anything else.

Cheers,
David.

[Vetal]

You won't like this advice, but...

1. Buy a new camcorder - preferably an HD one
2. Use more light

One or the other - preferably both.

If it's your child: I know it's an expensive time already, but they're only this young once!

Your camcorder has a typical annoying habit with consumer DV of hunting for the focus in low light, causing it to go really blurry every few seconds. It's really annoying, and there's no fix for it in AVIsynth or anything else.

Cheers,
David.


Why I won't like it? :)
Honestly, it is what exactly on my mind.

And it exactly the reason why I decided to move from Xvid + AC3 = > Avi to x264 + AAC world. To try new formats and tools. Now is just long justification and reasoning period :)

I had a lot of postings for the last 2 weeks and learn a lot of new stuff about HD. And you are very close in your advise.

In particular I'm thinking about Canon HG10. Not sure yet how it works in low light. As of reviews, it is not too good on dynamic changing scenes, has some problems on 24p export and has operating click recorded. Going to grab this toy soon to evaluate. And even more, to convince my wife, but this is ultimate task I'm gradually working on:)

Would be interested in useful links about cams. And opinions if it safe to buy it on eBay. I hope, I'm not too much off-topic.

Thanks

[Blue_MiSfit]

HG10 is nice, basically HV20 with a hard drive instead of a MiniDV deck.

I would recommend it, and definitely shoot in 60i for motion fluidity.

~MiSfit

[2Bdecided]

If buying on eBay, check that the warranty is valid.

If you want to convince your wife, try this (all examples bob deinterlaced, and resized downwards to square pixels)...

SD (http://www.david.robinson.org/pics/sd.jpg)
SD on an HDTV (http://www.david.robinson.org/pics/sdhd.jpg)
HD (http://www.david.robinson.org/pics/hd.jpg)

FWIW my backup is the original miniDV tapes. HDDs will work out cheaper soon, and faster already - but miniDV tapes can be thrown in a draw for ten years and still work. I have no confidence that multiple HDDs will survive that long without managed back up.

Cheers,
David.

[Vetal]

FWIW my backup is the original miniDV tapes. HDDs will work out cheaper soon, and faster already - but miniDV tapes can be thrown in a draw for ten years and still work. I have no confidence that multiple HDDs will survive that long without managed back up.

Cheers,
David.

For me, major drawback of the tape vs HDD is occasional frame drop-offs. So you never sure.

BTW, wich AR you use (or used) for DV, 8:9 or 10:11 ? I posted my question there ( http://forum.doom9.org/showthread.php?p=1099457#post1099457 ) , but no reply.

Thank you

[MadRat]

I notice you also have some color splotches (like on the baby's face) there that remind me of rainbowing. You might want to try an experiment with tritical's tcomb to see if that clears up. You'll have to tweak the settings but the documentation makes it really easy.

[2Bdecided]

For me, major drawback of the tape vs HDD is occasional frame drop-offs. So you never sure.You're right!

BTW, wich AR you use (or used) for DV, 8:9 or 10:11 ? I posted my question there ( http://forum.doom9.org/showthread.php?p=1099457#post1099457 ) , but no reply.I've answered there now, though others probably have more experience.

Cheers,
David.

[Vetal]

I notice you also have some color splotches (like on the baby's face) there that remind me of rainbowing. You might want to try an experiment with tritical's tcomb to see if that clears up. You'll have to tweak the settings but the documentation makes it really easy.

I tried it, it really helps!

But what a physical nature of the low-light rainbowing? I thought it kind of low-frequency noise? Do you have any links to read about it? Just want to find out, if HD will be free of it.