Is Intel really superior to AMD for video editing?

[poisondeathray]

Perhaps ffmpeg's method of computing PSNR is poorly implemented, that third picture looks like garbage. How did you create it? Is it from an I, P, or B frame?

One thing to consider, if your source is all I frames and your encode is of varying length gop chances are the frame that correlates with the time stamp may now be something other than an I frame and since P and B frames are usually encoded with a higher quantizer you are likely to get poor samples.

I don't see how that's possible, because this is a single frame lossless video, I frame. The "video" is 1 frame long.

I tested lossless Ut video RGB version of it, the PNG was just for demonstration purposes. I can upload the "video" if you want, but it' s just as easy to take the PNG and create a lossless 1 frame RGB video of your choice

Feel free to plug it into another test bed like MSU, avisynth , Elecard, etc...That's the reason why I resized it smaller so you can test freely available tools and compare in case one tool is not measuring correctly. They will give slightly different values, but all will basically say the same thing (ie. the trend will be the same) . MSU gave 23.65 to the degraded version, still supposedly "higher quality" than "2.avi" which achieved 22.27 accoring to MSU for PSNR-R (you can repeat for PSNR-G, and B as well but the trend will be the same)

Did you know why this is happening? I purposely degraded it from "1.avi", to show what 22-25 dB typical results would look like. This is degraded, without the 1px shift. This puts into perspective how damaging a 1 pixel shift is in terms of PSNR dB values. This is an example of how PSNR measure doesn't always correlate highly with human perception of "quality" or similarity. (Of course, there are many examples where it does correlate fairly well)

I hope it makes more sense to you why caution is advised when using PSNR metrics.

Do you want to see other examples, maybe more common or typical?

[sophisticles]

Perhaps ffmpeg's method of computing PSNR is poorly implemented, that third picture looks like garbage. How did you create it? Is it from an I, P, or B frame?

One thing to consider, if your source is all I frames and your encode is of varying length gop chances are the frame that correlates with the time stamp may now be something other than an I frame and since P and B frames are usually encoded with a higher quantizer you are likely to get poor samples.

I don't see how that's possible, because this is a single frame lossless video, I frame. The "video" is 1 frame long.

I tested lossless Ut video RGB version of it, the PNG was just for demonstration purposes. I can upload the "video" if you want, but it' s just as easy to take the PNG and create a lossless 1 frame RGB video of your choice

Feel free to plug it into another test bed like MSU, avisynth , Elecard, etc...That's the reason why I resized it smaller so you can test freely available tools and compare in case one tool is not measuring correctly. They will give slightly different values, but all will basically say the same thing (ie. the trend will be the same) . MSU gave 23.65 to the degraded version, still supposedly "higher quality" than "2.avi" which achieved 22.27 accoring to MSU for PSNR-R (you can repeat for PSNR-G, and B as well but the trend will be the same)

Did you know why this is happening? I purposely degraded it from "1.avi", to show what 22-25 dB typical results would look like. This is degraded, without the 1px shift. This puts into perspective how damaging a 1 pixel shift is in terms of PSNR dB values. This is an example of how PSNR measure doesn't always correlate highly with human perception of "quality" or similarity. (Of course, there are many examples where it does correlate fairly well)

I hope it makes more sense to you why caution is advised when using PSNR metrics.

Do you want to see other examples, maybe more common or typical?

I would like to continue this discussion but what do you say we start another thread, we have derailed this one enough as it is. And yes, I would greatly appreciate a more typical example, something more than a single frame.

I have been doing a ton of encoding tests trying to configure x264 to produce the highest quality encode as measured by both PSNR and SSIM, i.e. trying to get a value of 50db for PSNR and as close to 1 for SSIM and thus far quality, but my tests, the parameters that result in higher values for both also result in the more accurate encode, though perhaps not necessarily the better looking encode.

Perhaps it's psychological but I prefer the more accurate encode as measured by both PSNR and SSIM than the one that may look better in certain parts of the movie.

I intend to write a guide on how to do it if I can get PSNR to hit that magic number 50db and SSIM to be almost 1.

[poisondeathray]

I would like to continue this discussion but what do you say we start another thread, we have derailed this one enough as it is. And yes, I would greatly appreciate a more typical example, something more than a single frame.

I have been doing a ton of encoding tests trying to configure x264 to produce the highest quality encode as measured by both PSNR and SSIM, i.e. trying to get a value of 50db for PSNR and as close to 1 for SSIM and thus far quality, but my tests, the parameters that result in higher values for both also result in the more accurate encode, though perhaps not necessarily the better looking encode.

Perhaps it's psychological but I prefer the more accurate encode as measured by both PSNR and SSIM than the one that may look better in certain parts of the movie.

I intend to write a guide on how to do it if I can get PSNR to hit that magic number 50db and SSIM to be almost 1.

You should PM a mod, since it was you that derailed the thread

But that's one of the problematic parts of it - you're assuming that a higher PSNR or SSIM score is more "accurate" or more resembles the source video. Is it really ? Clearly not in the example above, and in many other cases. Of course, it's fine to prefer one or the other . Of course PSNR and SSIM are still used and can be useful. Also, when you get extremes , like high PSNR dB's , it becomes difficult with human eye to determine what is "closer" to source or "higher quality". Similarly abysmal quality encodes are hard to distinguish what is worse.

You have to look at single frames initially, to learn where PSNR and SSIM typically "fail" or give poor results that don't match with human perception. A lot of it you will learn by reviewing hundreds of tests. You will develop a "feel" for it. You can't do that with a single average PSNR number. You need to look at per frame values. You will develop your own conclusions, but I can tell you from experience that PSNR tends to over reward low noise, low detail encodes. Many encoders are tuned for PSNR, tend to have "plasticky" looking over smoothed results - similar to as if a denoiser was used. Common example is mainconcept AVC. In fact , if you use --tune PSNR for x264, psy and AQ are disabled, and the encode looks more similar to an mainconcept AVC encode.

Once you know how SSIM and PSNR typically "behave" . Its quite easy to "trick" or optimize for them with either encoding settings or filter preprocessing. You can trick average joe public and say "xyz" software has higher PSNR values. Common example of this is Video Enhancer - they post PSNR scores of their "upscaling", but some upscalers like NNEDI isn't corrected for the shift (using cshift), so is "misaligned" and scores much lower than it should have (it's also not enhanced with other filters like sharpening, but that' s another discussion)

What good is the 'average' PSNR? One number. Not very useful. For one thing, there are many different ways to calculate average PSNR. In YUV, "Y" is given higher weight usually, because human visual system is more sensitive to light/dark then color. Similarly in RGB, G channel is typically given higher weight. But what weight? Different formulas use different weighting. Another source of contention. They are useful for indicating problem areas , sections of encodes that might have problems (e.g you see a drop in the graph). Another example - how about a short feature, with exceptionally long titles and credits that were incredibly high quality, but the feature itself was crappy in terms of encoding quality. A single mean average for PSNR or SSIM will not be very representative of the arguably important parts and it will be a skewed representation. A single average PSNR or SSIM value is not so useful IMO

I dislike using x264 internal PSNR, SSIM, just "because." It's an internal metric and I avoid things that might be biased, or at least put less weight on them