The optimality of MB-tree.

[IgorC]

I would glad if x264 developers could answer my question.

Speaking basically, mb-tree moves the bits from complex parts of the scene (high motion - high temporal frequencies) to less complex (low temporal frequencies). Overall it's better in most of scenarios. Here is no doubt.

Some papers indicate that humans perception has maximum response on the middle (more precisely: low-middle) frequency range.

From this paper http://compression.ru/video/quality_measure/vqm.pdf

http://img209.imageshack.us/img209/2937/vqm.png
I have encoded some video and seen situation when background have exceedingly good quality while some middle frequencies content (people slow walking) weren't enough well preserved and all my concentration was directed into people. While it still makes a good job on lower quality of enough high motion where the details are hard to catch. I'm aware of that qcomp can lower MB-tree strength but it's actually lowers MB-tree positive impact as well.

Of course I'm not the expert to know how exactly MB-tree works and maybe strongly wrong.

Does it make any sense?


SSIM is a good metric but does it aware of what part of scene the human is watching at the moment to? (background or motion)

[cyberbeing]

You may want to compare the results of x264 MB-Tree vs Source and x264 without MB-Tree vs Source in the VQM metric with a video you think x264 is showing the issue you describe:
http://compression.ru/video/quality_measure/vqmt_download_en.html#start

Doing so should partially answer your question. Do you see a large change in the VQM metric without much change in SSIM?

[IgorC]

Sorry, I can't run tests in this moment.
Can you explain the idea of your previous post?

[Dark Shikari]

Part of the problem is that "where people look" depends on how you compress the video. If you reduce quality in an area of the video "because people don't look there", people will look there because the quality is noticeably worse there.

If you're interested in psy-tweaking MB-tree, feel free to drop by #x264dev; the VFR MB-tree code I added a few weeks ago should make this much easier.

[IgorC]

Dark Shikari,

I'm not sure if my statement was clear enough.
So, mb-tree prioritizes low frequencies (LF) comparing to middle (MF) and high ones (HF):
mb-tree bitrate allocation : LF > MF > HF . Is that right?
VQM suggestion : MF > LF > HF. (the bell curve)
The question is will be it more preferable visually.

If you reduce quality in an area of the video "because people don't look there", people will look there because the quality is noticeably worse there.

Yes, but it all depends on how much quality is increased or decreased.

[Dark Shikari]

If you have an actual VQM function, e.g. a C function that takes as input a time period (in seconds) and returns a weight based on that, I could run a blind test on that with MB-tree.

[IgorC]

Of course, I don't have it. :(