Avisynth: Convolution3d on optimal script?
 

Hello,

Is it a good idea if I add convolution3d to the optimal script ? It lowered my encoded file thus achieving a little higher in CQ...

or not? Thus it makes the output video more blurry?

My source is VHS, captured through DV cam passthrough, and using 352x480 kvcd resolution.

Thanks
dredj

In my experience, when adding it to the avisynth 2.5 script, convolution3d seems to increase the encode time about 12% and decrease the file size from 2% to 10% depending on source video. If you need a smaller file size try it. Image quality seems about the same with and without it (sometimes a bit better and sometimes a bit worse).

Thanks peter1234, yeah I too can't notice any differences, only the size of file...anyway, I might as well add it in order for me to increase the CQ since it lowers my file.

Also just to add this, I am also using deinterlacer...and base of the following deinterlace filter:

DGBob
SmoothDeinterlace
FieldDeinterlace

FieldDeinterlace seems to be promising..it lowers my filesize and encode time. (the test that i did for the 3 deinterlace filter does not include convolution3d btw).

Thanks,
dredj

You should also give "TomsMoComp" a try:
http://mywebpages.comcast.net/trbarry/TomsMoComp.zip

For me, it gives a little less "Jaggies" on edges than "FieldDeinterlace",
while also improving on compression.


*******************************
The Devil`s always.....in the Details!

Thanks I'll give it a try...follow up question, what is the best searcheffort value that you are using? and should I turn on vertical filter?

Regards,
Dredj

Re: convolution3d on optimal script
 

Nope ;)

-kwag

The best deinterlacer I have found for file size reduction is the TMPGEnc deinterlace (double) advanced filter. For one video that had bad interlacing, generating a deinterlaced AVI using TMPGEnc deinterlace (double) and then using that AVI as the source for avisynth script reduced the mv2 file size an additional 22% compared to convolution3d alone when using Kwag's motion adaptive script. Shows how much extra file size can be added by interlace noise. I haven't found any deinterlacers that work as well that can be used within the avisynth script. But I sure wish someone would make an avisynth script version of the TMPGEnc (double) deinterlacer. I have tried Decomb, SmoothDeinterlacer, and FieldDeinterlace. FieldDeinterlace(full=true, chroma=true) seemed to work best but had some skip problems with Kwag's script at missing fields and only reduced file size about 10%. If you find a good deinterlacer that can be used in the avisynth script please let me know.

Kwag,
Do you think a 12% increase in encode time is worth a 2% reduction is file size? I suspect most people are not that worried about a few percent file size reduction and would rather have a quicker encode time, so I would recommend not adding it to script.

No :mrgreen: The current script does a great job as it is. Not to mention that when you use C3D, you loose some of the "natural" look of the picture. I'm very pleased with the script as it is, and the only things that will (should?) be added should be interlaced related stuff. I believe FredThompson is working on just that, so we'll barrow intelaced related stuff from him ;)
Unless some new filter comes around that drops file size considerably, I don't see any need to add more bulk to the script.
I like the concept of K.I.S.S :mrgreen:

-kwag

Thanks Kwag, i'll remove it then.

For the deinterlace, I tried vico1 suggestion which is tomsmocomp and seems to be okay execept i found some vertical line on some of the clip, this is probably i turn on the verticalfilter paramater of tomsmocomp, maybe i should turn it off....anyway i tried viewing it to my TV, and i will say there's not much difference i saw among the deinterlace filter that i use.

Peter123, If i understand you correctly, you said u use tmpgenc deinterlace(double) to deinterlace an input avi file and output it to avi file also, am i right? Can you tell me how did you accomplish that? I want to try it?

Thanks....

dredj

as an addition, can we just use tmpgenc deinterlace filter while encoding to final mpeg instead of avi to avi then to mpeg?

Just asking....anyway let me try it also....

dredj

dredj,
File needs to be deinterlaced before going through script. Using TMPGEnc at the end will not give the same results. To generate the AVI, setup everything in TMPGEnc as if encoding, with deintrlace (double) checked on advance tab, but use the output to file option on the file menu to generate AVI instead of clicking start button.

PS - When I tried TomsMoComp filter it increased encode time several hundred percent and didn't seem to give any better output than just deinterlacing. I wonder if I was doing something wrong.

Peter123, also have you consider Double(Adaptation) instead of just Double? seems like using Double blurrs the picture more as oppose to Double(adaptation).

anyway back to my question to u..how did u accomplish and avi to avi through tmpgenc?

Thanks,
dredj

dredj,
To generate the AVI, setup everything in TMPGEnc as if encoding, with deintrlace (double) checked on advance tab, but use the output to file option on the file menu to generate AVI instead of clicking start button.

Thanks for the info Peter123, let me try that one now....in the meantime here is the comparison for all the deinterlace that I use in terms of avg encode time and sample output filesize, my PC is P3 FSB133 800MHZ 512MB.

Using TOK 0.0.5.3 default setting and CQ 60

Deinterlace filter --> Average encoded time (min) --> File size:

DGBob --> 8min --> 12,394,348
SmoothDeinterlace --> 8min --> 12,342,320
FieldDeinterlace --> 5 min --> 11,834,150
TomsMoComp --> 5 min --> 11,601,925
TMPGenc Deinterlace (Double adaptation) --> 10min --> 12,112,221
TMPGenc Deinterlace (Double) --> 5 min --> 11, 494,792

Now let me try your suggestion peter123.

Thanks,
dredj

The bluring is what reduces the file size. It is motion adaptive filtering for moving edges. Removing the noisy sharp moving edges reduces the data bits required to encode the video. This approach only blurs the moving edges, whereas Kwag's filter blurs the entire frame.

okay additional question peter123, did u use no compression or u use huffyuv ? what config settings did u use when you use huffyuv, is it:

YUY2 compression method --> Predict median(best)
RGB Compression method --> <-- Convert to YUY2
Field Threshold 480
and checked "Enable full size output-buffer"

?

Thanks,
Dredj

dredj,
I used no compression, generated RGB avi for input to avisynth, then ConvertToYV12 after reading into avisynth script.

dredj,
This is getting rather complex. I simply wanted to point out that interlace noise can greatly increase file size. It looks like you have already reduced your file size by about 10% simply be using deinterlace (double) when you did your TMPGEnc encoding. Unless you have a lot of interlace noise I don't think it is worth the effort to do more than that.

with help from kwag, I've put together a version of the filter which is for interlaced source. I'll release the full set of scripts in the next few days. They need a little more work.

If you must deinterlace or want to decomb, try Donald Graft's new beta available in the doom9 forums. If your source is interlaced, i.e., pure video and MPEG2 is an option, consider NOT deinterlacing because you really don't have 25/30 fps, you have 50/60 fps of half-height pictures which don't match on a perfet 1:1 correlation. Deinterlacing will always introduce a LOT of distortion. It might not be visible to you, but it's there. You could play with bobbing but still, 50% of each frame is a "best guess", same as deinterlacing.

Decombing to get progressive film source is another matter.

Having said that, if your source isn't complex with a lot of motion, maybe you'll be happy with de-=interlaced results.

Try it on something like WWE wrestling opening credits/"fights" and you'll see lots of distortion. Flash photography doesn't play well with deinterlacing because it's one-field, not one-frame...well...multiple different flashes could do consecutive fields.