Video quality metric

[koliva]

I know that PSNR is most widely used objective video quality metric but I am not sure that it is used in H.264. What metric is used in H.264?

Thanks.

[LoRd_MuldeR]

I think H.264 doesn't use any metric per definition. Developers may use any metric they find useful to tune their H.264 encoder. x264 calculates both, PSNR and SSIM.

But you have to notice that these metrics are useless to judge the overall quality of a video, especially when Psy-Optimizations are used!

[koliva]

I think H.264 doesn't use any metric per definition. Developers may use any metric they find useful to tune their H.264 encoder. x264 calculates both, PSNR and SSIM.

But you have to notice that these metrics are useless to judge the overall quality of a video, especially when Psy-Optimizations are used!

Your answer cause a new question. Why are they useless? What is Psy-Optimization?

[LoRd_MuldeR]

Your answer cause a new question. Why are they useless? What is Psy-Optimization?

They are useless because numbers can't represent the visual quality. The human eye doesn't follow any metrics ;)

Consequently metrics can only measure the difference between the encoded video and the original, but they can NOT measure the perceived (subjective) quality!

Psy-Optimizations are "psycho-visual" optimizations that try make the video look "better" for the human viewer. However they hurt metrics like PSNR and SSIM.

It seems that metrics prefer a "smooth" video, whereas people usually prefer a more "sharp" and detailed/grainy video, even if that video differs more from the original.

Anyway, some aspects of an encoder can only be tuned by using objective metrics (PSNR/SSIM). Just don't use them to judge overall quality...

[koliva]

They are useless because numbers can't represent the visual quality. The human eye doesn't follow any metrics ;)

Consequently metrics can only measure the difference between the encoded video and the original, but they can NOT measure the perceived (subjective) quality!

Psy-Optimizations are "psycho-visual" optimizations that try make the video look "better" for the human viewer. However they hurt metrics like PSNR and SSIM.

It seems that metrics prefer a "smooth" video, whereas people usually prefer a more "sharp" and detailed/grainy video, even if that video differs more from the original.

Anyway, some aspects of an encoder can only be tuned by using objective metrics (PSNR/SSIM). Just don't use them to judge overall quality...


Thanks for your detailed answer. I wrote a code and I can compress image sequences using motion vector like H.264. Somehow, I need to compare my results with the original ones. If I use PSNR, it gives around 35 but when I look at my result images, I can easily see the errors. I thought that there is another metrice that can give me my error more objectively. I will check SSIM. Maybe it is better than PSNR.

[audyovydeo]

Thanks for your detailed answer. I wrote a code and I can compress image sequences using motion vector like H.264. Somehow, I need to compare my results with the original ones. If I use PSNR, it gives around 35 but when I look at my result images, I can easily see the errors. I thought that there is another metrice that can give me my error more objectively. I will check SSIM. Maybe it is better than PSNR.


Koliva,
the pros and cons of metrics as applied to x/H.264have been debated extensively in this forum. I suggest you search around and sieve through the pros/cons.


cheers
a/v

[benwaggoner]

...and one big issue with "classic" metrics is they assume a perceptually uniform gamma curve (so going from 16 to 17 is the same visible difference in brightness as going from 216 to 217). But with LCD displays, we aren't perceptually uniform, so banding/blocking in the low luma range is a lot more obvious than it is in the rest of the range.

[CruNcher]

absolutely agree ben i released a comparison one time here of this visual problem i always try to optimize that layer by amplifying it in my preprocessing task and get it smoother out after all the compression quantization still hurts this heavy but with more enhancements since then in X264 it's much better :)
http://forum.doom9.org/showthread.php?p=1114782#post1114782
Though im still very shocked seeing what should be Professional H.264 Encodings in Main Profile lose vs WMV MP@HL :P in terms of sharpness it's always the same most H.264 encoder prefer smoothing and even X264 did for a long time if Dark Shikari wouldn't have joined in im not sure if X264 would stand where it does now Visually as most people are only interested in enhancing pure mathematical efficiency :). But still this doesn't yet reached most content creators and so still it will go on that most WMV Encodings will look visually more appealing to many then H.264 encodes in this area and that is really sad :(

[Gabriel_Bouvigne]

But with LCD displays, we aren't perceptually uniform, so banding/blocking in the low luma range is a lot more obvious than it is in the rest of the range.
And it's even worse with badly calibrated LDC panels