VideoHelp Forum
+ Reply to Thread
Page 6 of 6
FirstFirst ... 4 5 6
Results 151 to 168 of 168
Thread
  1. Banned
    Join Date
    Nov 2005
    Location
    United States
    Search Comp PM
    Originally Posted by poisondeathray View Post
    --tune ssim if you 're measuring ssim , --tune psnr if you're measuring psnr

    sure it will yield higher metrics, but usually worse subjective image quality
    oh, what's the matter? a bit upset because x264 gets it's ass kicked when measuring quality with an objective metric like SSIM?

    btw, DS has claimed that mb-tree increases SSIM by up to 70%, so what would the scores be if mb-tree were disabled?

    LOL@x264.
    Quote Quote  
  2. Banned
    Join Date
    Feb 2013
    Search PM
    DivX265 doesn't have those so I figured I'd keep it as even as possible. I can re-run the test tomorrow though.

    P.S.
    Your fix not only seems to have taken care of my regular AviSynth 2.6a folder but now hybrid's AviSynth functions wotk for me, without having to touch hybrid's folder. Might just be a coincidence
    Quote Quote  
  3. Originally Posted by deadrats View Post
    Originally Posted by poisondeathray View Post
    --tune ssim if you 're measuring ssim , --tune psnr if you're measuring psnr

    sure it will yield higher metrics, but usually worse subjective image quality
    oh, what's the matter? a bit upset because x264 gets it's ass kicked when measuring quality with an objective metric like SSIM?

    btw, DS has claimed that mb-tree increases SSIM by up to 70%, so what would the scores be if mb-tree were disabled?

    LOL@x264.



    Why would I be upset? LOL . I just make observations. I'm just interested in the truth

    The truth is (at least with x264, I haven't used x265 enough to see if this holds true for x265) - Typically you get higher values by tuning SSIM and PSNR, occasionally this isn't the case, but try it out yourself . Obviously you know I've done a gazillion tests, trust me on this

    These metrics are of limited value anyway. Take actual look at the image quality

    And there is that "by up to xxx%" claim again. I'd take those marketing words with a grain of salt

    Why are you so emotional ?
    Quote Quote  
  4. Originally Posted by gonca View Post
    DivX265 doesn't have those so I figured I'd keep it as even as possible. I can re-run the test tomorrow though.
    DivX265 doesn't, but their encoder, like all Mainconcept varieties are generally tuned for PSNR at default. Their engineers optimize for this, just like many other codec developers. You can tell by the encodes in the other threads, the hallmark is smoothing and lack of detail, esp. in shadow areas. Things like psy, AQ typically reduce SSIM and PSNR values, but subjectively they do make image look better (not always, but most of time). Actually if you turn down AQ, psy, x264 begins to look a lot like Mainconcept encodes

    Objective metrics have their place, but they are only limited in value. I won't discuss it here because pros/cons are discussed extensively elsewhere

    Tests are usually designed run with an "endpoint" in mind. So if the endpoint is SSIM, you usually encode for SSIM . You don't handicap an encoder just to level the playing field, you show the encoder in the best possible light
    Quote Quote  
  5. Banned
    Join Date
    Feb 2013
    Search PM
    The only true test of quality are your own eyes and taste. One man's trash is another man's riches.
    Quote Quote  
  6. Banned
    Join Date
    Nov 2005
    Location
    United States
    Search Comp PM
    Originally Posted by poisondeathray View Post
    And there is that "by up to xxx%" claim again. I'd take those marketing words with a grain of salt
    i take anything DS says with a pound of salt (which he should go do) and some pepper for good measure. but they are not "marketing words" they are DS' quotes when he first introduced mb-tree over at doom9.
    Quote Quote  
  7. Banned
    Join Date
    Nov 2005
    Location
    United States
    Search Comp PM
    Originally Posted by poisondeathray View Post
    DivX265 doesn't, but their encoder, like all Mainconcept varieties are generally tuned for PSNR at default. Their engineers optimize for this, just like many other codec developers. You can tell by the encodes in the other threads, the hallmark is smoothing and lack of detail, esp. in shadow areas. Things like psy, AQ typically reduce SSIM and PSNR values, but subjectively they do make image look better (not always, but most of time). Actually if you turn down AQ, psy, x264 begins to look a lot like Mainconcept encodes

    Objective metrics have their place, but they are only limited in value. I won't discuss it here because pros/cons are discussed extensively elsewhere

    Tests are usually designed run with an "endpoint" in mind. So if the endpoint is SSIM, you usually encode for SSIM . You don't handicap an encoder just to level the playing field, you show the encoder in the best possible light
    for the love of Gandhi! main concept's h264 has AQ and in fact they have 3 or 4 different settings, but only some apps expose them all. also the divx hevc encoder does have AQ, that's what the presets refer to, which may explain why SSIM was lower for the slower settings than the faster with divx.

    by the way, despite the fact that DS spent 10+ years discrediting mathematical measurements like SSIM and PSNR, the fact remains he has admitted that they are an accurate method of testing changes within an encoder, i.e. measuring the impact of certain settings and he has used it when convenient to promote his half-assed encoder.

    more importantly, the MSU comparisons that his supporters always point to as proof that x264 is the best encoder all use math metrics as their method of determining a winner, look at their reports it's full of average and peak SSIM and PSNR readings.
    Quote Quote  
  8. Originally Posted by deadrats View Post


    for the love of Gandhi! main concept's h264 has AQ and in fact they have 3 or 4 different settings, but only some apps expose them all. also the divx hevc encoder does have AQ, that's what the presets refer to, which may explain why SSIM was lower for the slower settings than the faster with divx.

    Yes, but they don't work very well. I wrote about this many years ago, posted comparisons - I think you particpated in that thread.


    by the way, despite the fact that DS spent 10+ years discrediting mathematical measurements like SSIM and PSNR, the fact remains he has admitted that they are an accurate method of testing changes within an encoder, i.e. measuring the impact of certain settings and he has used it when convenient to promote his half-assed encoder.
    They have their place, because it's tough to review 1000's of encodes frame by frame


    more importantly, the MSU comparisons that his supporters always point to as proof that x264 is the best encoder all use math metrics as their method of determining a winner, look at their reports it's full of average and peak SSIM and PSNR readings.
    Mainly SSIM. But you're right about that point - it's just "part" of the whole picture. Definitely not "proof" . You have to look at other measurements, other tests as well . But when you take all those other measurements, subjective measures, there is still only 1 clear winner . You're just too biased to see it. Seriously Mountains of proof!




    Originally Posted by deadrats View Post
    Originally Posted by poisondeathray View Post
    And there is that "by up to xxx%" claim again. I'd take those marketing words with a grain of salt
    i take anything DS says with a pound of salt (which he should go do) and some pepper for good measure. but they are not "marketing words" they are DS' quotes when he first introduced mb-tree over at doom9.
    Yeah, so he picks one outlier that happens to be x% and that makes the statement logically true . Marketing 101
    Quote Quote  
  9. Sr Manager Broadcast
    Join Date
    Jun 2014
    Location
    Bangalore, India
    Search PM
    Hi @ Jagabo

    Though PSNR values might not give impression on quality, the range of PSNR itself tells on the probable quality of video output. If not PSNR what are the metric used to compare. What are the BD rate gains in case of DiVXh265 as compared to HM. hence the suggestion was to have some comparison metric.
    Quote Quote  
  10. @deadrats: LOL, playing the stupid brat again and getting all folks rattled *congratulation*
    It's sad that the thread get's derailed with all the name calling and bs,..
    Quote Quote  
  11. Originally Posted by Selur View Post
    @deadrats: LOL, playing the stupid brat again and getting all folks rattled *congratulation*
    It's sad that the thread get's derailed with all the name calling and bs,..
    There's one in every forum, and here it's deadrats.
    After claiming in the other thread the Divx h264/5 encoders don't have a deblocking filter, I'd have hoped he'd refrain from posting in any of the h264/265 threads for a little while.... till most of us have forgotten he doesn't know what he's talking about, but apparently not......

    You'd expect a grown-up to return to at least post "I was wrong", but for reasons I don't understand being an adult is hard for some, even behind the anonymity of a username and the internet.
    Quote Quote  
  12. Banned
    Join Date
    Feb 2013
    Search PM
    ssim and psnr results
    CRF 18 preset medium
    Image Attached Thumbnails CRF 18 medium preset.pdf  

    Last edited by gonca; 8th Jul 2014 at 19:30.
    Quote Quote  
  13. Member
    Join Date
    Aug 2013
    Location
    Central Germany
    Search PM
    A metric is a mathematical algorithm which will never accurately measure subjective impressions of all people. No machine can measure if "it looks better". So please, stop overestimating the meaning of PSNR, SSIM, and whatever metric will be developed in the future. Quality is a fuzzy term, and even the meaning of an ABX test of thousands of probands is limited.
    Quote Quote  
  14. Banned
    Join Date
    Feb 2013
    Search PM
    Originally Posted by LigH.de View Post
    A metric is a mathematical algorithm which will never accurately measure subjective impressions of all people. No machine can measure if "it looks better". So please, stop overestimating the meaning of PSNR, SSIM, and whatever metric will be developed in the future. Quality is a fuzzy term, and even the meaning of an ABX test of thousands of probands is limited.
    I am aware of that but if you look at previous posts you will see that I was asked to run these tests
    Quote Quote  
  15. Member
    Join Date
    Aug 2013
    Location
    Central Germany
    Search PM
    I don't mind you running tests and reporting values.

    I am just afraid of generalizations like "this number is smaller, therefore that quality is better". That may often be true if the difference is remarkable (debating about permille differences is meaningless). But there are academic samples how to fool metrics (e.g. PSNR is not as sensitive as human perception to impulsive "Salt & Pepper" noise, but very sensitive to already slight contrast changes or blur).

    A better metric value only means "a better metric value", but not certainly "better quality preservation for the average human recognition".
    Quote Quote  
  16. Banned
    Join Date
    Feb 2013
    Search PM
    I'm only running the tests because I was asked to, and I offer no opinions on the results.
    Personally, I believe that the only metric that counts is the one that uses your eyesight and viewing equipment.
    Quote Quote  
  17. Member
    Join Date
    Aug 2013
    Location
    Central Germany
    Search PM
    That would probably be an ABX test with as many participants as possible. Like Roberto Amorim once did for audio formats, with support by the HydrogenAudio forum users.
    Quote Quote  
  18. Originally Posted by deadrats View Post
    Originally Posted by Mephesto View Post
    I leave for a few months
    yeah, those parole violations can be very inconvenient. are you even supposed to be using a computer? maybe you should just stick with the community service for a while.
    Not only do you know I'm on parole but you know the exact crime I've done that restricts use of PCs upon release. It must be awesome to know everything about everyone.

    EDIT: About the debate about usefulness of SSIM. It is accurate to compare large disparities of scores but not small ones. An SSIM score of 0.99050 vs 0.99100 doesn't prove shít. But a 0.98500 vs 0.99000 shows a clear difference that a majority of the time is a lot better quality.
    In the million SSIM tests I've done and tests I've seen others do, I am yet to see an example of a video getting an extremely lower SSIM score yet being better quality.

    Yes, I've seen scores as low as 0.89000 where the quality wasn't so bad and ones as high as 0.98500 that looked like garbage but never a large disparity of the same video without the quality perfectly correlating to it.
    Consistency is the key and not comparing apples and oranges with an "objective" metric and expecting accuracy.

    x265, x264, Xvid, MPEG-2 are all refined versions of the same DCT algorithm so they are easy to compare with the same metric accurately.

    Try comparing wavelet and DCT at the same bitrate like JPEG and JPEG 2000. It'll be hard to tell which ones are better even when using your eyes. In many cases the wavelet photo won't look necessarily better, just different. You try to decide which artifacts are more annoying and which details are more important to retain. It's a mindfück.
    Last edited by Mephesto; 12th Jul 2014 at 11:37.
    Quote Quote  
Visit our sponsor! Try DVDFab and backup Blu-rays!