SIF1 Video Codec

[Neiromaster]

The SIF1 project was launched in the autumn of 2000, with the major goal of finding new approaches for video compression. We developed a new class of algorithms for image compression. Its working title is "SIF-conversion". This class of algorithms can be realized in a number of ways, and the total efficiency of the resulting compression engine depends on the method of realization chosen. During several years, we worked on different compression engines based on SIF conversion. The sixth compression core was found to be the best basis for realizing the new codec. It should be noted that it could be further improved since the developed engine represents a compromise between the compression efficiency obtained and the time spent on the engine development.

History of versions

0.45 Alpha4
1. Decoding speed has been increased over 1.6 times.
2. The total compression efficiency has been increased.
3. The level of some artifacts of image compression has been decreased.
4. The psychovisual model has been appreciably increased.

0.40 Alpha3
1. The algorithm has been improved and the efficiency of the operation of the Core6 analyzing part has been increased. The level of some artifacts on the image has been decreased.
2. The algorithm of the quantizer operation has been changed.
3. An adaptive algorithm of the bit rate control had been added.
4. The sharpness of the decoded image has been appreciably increased.
5. The psychovisual model has not been finally debugged. However, it operation is acceptable.
6. The total compression efficiency has been considerably increased.
7. This version is not compatible with the previous one, but all subsequent versions of the codec will decode the video compressed by the former version.

0.30 Alpha2
1. The entropy codec has been corrected and modified.
2. The level of some artifacts of image compression has been considerably decreased.
3. The functionality of the motion compensation engine has been improved.
4. An automatic arrangement of key frames has been added.
5. The compression efficiency has been appreciably increased.
6. Support of input color spaces RGB32, YUY2, and YV12 has been added.
7. This version is fully incompatible with the previous one.

0.25 Alpha1 - The first public version.

Further plans

1. To add new compression regimes in which the size of the output compressed file can be specified.
2. To improve the operation of the motion detection engine and to add the new higher quality regimes that are oriented at real-time compression.
3. To develop an option of using quarter-pixel motion compensation. In present version, standard half-pixel compensation is used.
4. To further improve the operation of the analyzing part of Core6.
5. To perform SSE2 and multiprocessor optimization of the current code.
6. To develop a new core (Core7) and SIF 2 codec on its basis.


http://mysif.ru/SIF1_dd_Eng.htm

[Sirber]

Is it VFW? Is it CLI? What container supports it? How is it compared to XviD?

[LoRd_MuldeR]

Originally Posted by Sirber

Is it VFW? Is it CLI? What container supports it?

http://img109.imageshack.us/img109/8009/sifvz6.gif
I guess the answers are "VfW" and "AVI"

// EDIT

DirectShow decoder filter for playback is included.

[Neiromaster]

Originally Posted by Sirber

What container supports it?

AVI, Matroska, Ogm

Originally Posted by Sirber

How is it compared to XviD?

Between x264 and XviD. Closer XviD.

[bratao]

Good,
REAL Good !!

Keep the good work, you maybe can create a true alternative to the mpeg products !

By the way, how the internal of this codec, its wavelet, macroblock... ??

[mod]

Any possibility to see it on Linux? VFW doesn't work well on it ^^

[Neiromaster]

Originally Posted by bratao

Good, REAL Good !!

Thanks.

Originally Posted by bratao

By the way, how the internal of this codec, its wavelet, macroblock... ??

It based on the adaptive block size overlapped motion compensation and set of overlapped subband transforms.
A species of transform determined local statistic of image.

It isn't wavelet, it isn't block transform, its SIF - absolutely new patentable algorithm.

[Neiromaster]

Originally Posted by mod

Any possibility to see it on Linux?

It will be possible in future.
I'm planning release decoding library on VP3-like license in future.

[pardus]

Originally Posted by Neiromaster

It will be possible in future.
I'm planning release decoding library on VP3-like license in future.

Why not add it to ffmpeg under L/GPL ? Maybe even dual license for commercial usage.

[bratao]

I´m checking some samples of your site and i like it.

The exemple that i see is the Matrix "metro" one. ITs 200 Kbits/s
I like the image is degradation of low bitrate , it isnt "block", it like "noise" the image .

In this such premature stage(no configuration, no 2 pass,...) , its a very exciting codec !

[KoD]

"L/GPL"

Just because those are buzzwords on everyone's lips it doesn't mean they're actually what you want.

[MfA]

As long as there are no patents involved there really isn't much use going with the original VP3 license over LGPL.

[GodofaGap]

You should compare both licenses first before making such a statement. VP3 grants almost complete freedom, as long as you don't use patented algorythms themselves to make something completely different. LGPL is far more restrictive.

[BeyondTheEyes]

Originally Posted by Neiromaster

Thanks.

It based on the adaptive block size overlapped motion compensation and set of overlapped subband transforms.
A species of transform determined local statistic of image.

It isn't wavelet, it isn't block transform, its SIF - absolutely new patentable algorithm.

- SIF stands for ?
- any preprocessing (e.g. deblocking) ?
- any adaptive frame-skipping ?
- any rate-distortion optimization ?
- any ...
- any chance http://mysif.ru/O_SIF.htm to be translated into english ?

- anyway, it looks (very) promising ! ;-)

PS : Are OBMC and M-band decompositions really that "new" ?

[Neiromaster]

Originally Posted by BeyondTheEyes

- SIF stands for ?

Now its just SIF

Originally Posted by BeyondTheEyes

- any preprocessing (e.g. deblocking) ?

It not use any pred-postprocessing technique and inloop filtering.
I'm working at otner path to increase quality.

Originally Posted by BeyondTheEyes

- any adaptive frame-skipping ?

It isn't in early plans.

Originally Posted by BeyondTheEyes

- any rate-distortion optimization ?

It stands in early development stady.

Originally Posted by BeyondTheEyes

- any chance http://mysif.ru/O_SIF.htm to be translated into english ?

Maybe.

Originally Posted by BeyondTheEyes

PS : Are OBMC and M-band decompositions really that "new" ?

1) I didn't say that OBMC is new.
2) M-band decompositions != subband transform
For example: Laplacian Pyramid and Steerable Pyramid is not M-band but its subband transforms.

[LoRd_MuldeR]

Will there be official specs available to the public, so developpers can implement their own encoders/decoders? I ask, because I doubt it will ever become popular without proper ffmpeg/lavc support...

[BeyondTheEyes]

Originally Posted by Neiromaster

For example: Laplacian Pyramid and Steerable Pyramid is not M-band but its subband transforms.

Would you "disclose" that SIF1 is using an overcomplete representation ?-) If, yes, may I kindly ask how you achieve to remove (or at least minimize) the underlying redundancy ?

BTW, I have just made a quick test with SIF1 (DVD source -> 640x272@25Hz, ~500 Kbps) and I must admit that the result is quite impressive (in terms of perceived quality, not PSNR, which may not be appropriate here to assess the quality of SIF1 in comparison with traditionnal hybrid DCT-based codecs, right ?)

[foxyshadis]

PSNR isn't an appropriate measurement for anything, except matching up individual blocks in the codec. At minimum use SSIM, at best use your eyes on tough scenes.

Originally Posted by LoRd_MuldeR

Will there be official specs available to the public, so developpers can implement their own encoders/decoders? I ask, because I doubt it will ever become popular without proper ffmpeg/lavc support...

Popularity among companies or academia generally has nothing to do with popularity among the open source crowd.

Subband transforms were somewhat popular in the late 80's and early 90's, but fizzled out as mpeg-4 made its appearance and delivered similar efficiency with better performance, and scaled up better. They're still useful at the bottom end for the same reason jpeg2k is much better than jpeg at the bottom end, but I'm interested that performance here seems relatively good even at higher birates.

Still, I was under the impression that pretty much anything remotely related to subband encoding had been long patented away, another reason they aren't so popular now. Or have they started to fall out of patent?

[LoRd_MuldeR]

Originally Posted by foxyshadis

Popularity among companies or academia generally has nothing to do with popularity among the open source crowd.

Yeah. But do you really think the companies are interesetd in a completely new standard like SIF1 when H.264/AVC and WMV/VC-1 are on their way ???
I doubt that, even if SIF1 had really good results...

But I think it could become popular very easy among the OpenSource crowd when it perfroms good and has a proper implementation in ffmpeg (or at least a free encoder/decoder lib)

[MfA]

Basically subband transforms are a subset of wavelet transforms, they predate wavelet theory.

Why are you ignoring lifting though? Fundamentally lossy transforms are just not acceptable anymore IMO. We have to learn from the past, MPEG4 and it's problems with varying DCT implementations specifically.