Page 2 of 2 FirstFirst 12
Results 31 to 38 of 38

Thread: SSIM / MSSIM vs MSE

  1. #31
    Member
    Join Date
    Apr 2012
    Location
    Stuttgart
    Posts
    448
    Thanks
    1
    Thanked 101 Times in 61 Posts
    Quote Originally Posted by cbloom View Post
    I'd love to see the power of Google used to generate a new large-scale human rating database of image distortions.
    Such databases already exist. LIVE, LIVE2, TIDE2013 (sp?) to name three. They are available from the corresponding authors. Distortions are (for LIVE and LIVE2) JPEG, JPEG 2000, gaussian blur, transmission loss (fading channel) and white noise.

  2. #32
    Member
    Join Date
    Jun 2015
    Location
    Switzerland
    Posts
    976
    Thanks
    266
    Thanked 350 Times in 221 Posts
    Quote Originally Posted by thorfdbg View Post
    Such databases already exist. LIVE, LIVE2, TIDE2013 (sp?) to name three. They are available from the corresponding authors. Distortions are (for LIVE and LIVE2) JPEG, JPEG 2000, gaussian blur, transmission loss (fading channel) and white noise.
    Do you mean TID2013 by TIDE2013, or are they two different corpora?

  3. #33
    Member
    Join Date
    Sep 2010
    Location
    US
    Posts
    126
    Thanks
    4
    Thanked 69 Times in 29 Posts
    Quote Originally Posted by thorfdbg View Post
    Such databases already exist. LIVE, LIVE2, TIDE2013 (sp?) to name three. They are available from the corresponding authors. Distortions are (for LIVE and LIVE2) JPEG, JPEG 2000, gaussian blur, transmission loss (fading channel) and white noise.
    In my work I used TID2008. (it was a while ago). I thought it was absolutely terrible, almost useless.

    1. Most of the distortions are way too large. They're way out of the real working range of compressors.

    2. Many of the distortions are things that compressors will never do or encounter in real use.

    3. They're missing lots of distortions that would be interesting, like variable downsampling of color planes, variable bit allocation of color planes, non-spatial uniform bit allocation of a compressor (local vs. global optimization problems), whether deblocking & other post-processing improves human visual quality, etc. etc.

    Of the 60+ distortion variants, only 4 or so are actually useful. Everything else skews the competition in a weird invalid way.

    The distortions in the test set need to be asking questions that compressors will see in the real world, they need to be within the reasonable working range of image quality (just slightly visible). Is it better to dequantize with randomness in the quantization interval? Is it better to rate allocate to preserve energy level? etc.

    I was pretty easily able to definitively beat any of the metrics from the literal in terms of fitting the TID2008 scores :

    http://cbloomrants.blogspot.com/2011...l-metrics.html

    but the more I worked on it the more I felt that was a bogus result without a better test set.

    Summary post :

    http://cbloomrants.blogspot.com/2012...ison-post.html
    Last edited by cbloom; 3rd August 2016 at 20:11.

  4. Thanks (2):

    Jyrki Alakuijala (16th April 2016),Turtle (16th April 2016)

  5. #34
    Member
    Join Date
    Sep 2010
    Location
    US
    Posts
    126
    Thanks
    4
    Thanked 69 Times in 29 Posts
    Quote Originally Posted by Jyrki Alakuijala View Post
    Would you consider open sourcing your solution?
    I'd be happy to, but it's rather too much of a mess, and I don't have time to clean it up. (speaking of which, I really need to finish my JPEG decoder!)

    The winner for me "MyDctDelta" is very simple, just an 8x8 DCT at every pixel location, with JPEG CSF, some variance masking, and AC category preservation terms.
    Last edited by cbloom; 3rd August 2016 at 20:11.

  6. #35
    Member
    Join Date
    Jun 2015
    Location
    Switzerland
    Posts
    976
    Thanks
    266
    Thanked 350 Times in 221 Posts
    Quote Originally Posted by cbloom View Post
    I'd be happy to, but it's rather too much of a mess, and I don't have time to clean it up. (speaking of which, I really need to finish my JPEG decoder!)

    The winner for me "MyDctDelta" is very simple, just an 8x8 DCT at every pixel location, with JPEG CSF, some variance masking, and AC category preservation terms.
    What are AC category preservation terms?

  7. #36
    Member
    Join Date
    Apr 2012
    Location
    Stuttgart
    Posts
    448
    Thanks
    1
    Thanked 101 Times in 61 Posts
    Quote Originally Posted by cbloom View Post
    In my work I used TID2008. (it was a while ago). I thought it was absolutely terrible, almost useless.
    It rather depends which questions you ask. For compression performance, you're typically interested in distortions at the visual threshold. For such distortions, this database is indeed not very useful. You're typically good with the known effects (CSF, visual masking) which are well-implemented in VDP. It's working pretty well in this domain. There is, however, another question, and that's the question of super-threshold distortion. It is also a relevant question, but it's much harder to evaluate and quantify, and it is less relevant for image compression applications. The TID and LIVE databases are more tuned for those applications. This said, I'm not very happy with the databases either. The images are too small, the quality is not very good, and it's not exactly the type of distortion I personally care for.

  8. #37
    Member
    Join Date
    Apr 2012
    Location
    Stuttgart
    Posts
    448
    Thanks
    1
    Thanked 101 Times in 61 Posts
    Quote Originally Posted by cbloom View Post
    The winner for me "MyDctDelta" is very simple, just an 8x8 DCT at every pixel location, with JPEG CSF, some variance masking, and AC category preservation terms.
    Well, I made a very similar image quality score years ago, essentially masking and a csf model. It also worked pretty fine, but it never gained much attention in the community. There's really nothing innovative about it. Anyhow, if there is some interest, I can certainly dig out the source. There's also a small paper about it.

  9. #38
    Member
    Join Date
    Jul 2016
    Location
    Russia
    Posts
    26
    Thanks
    17
    Thanked 10 Times in 7 Posts
    SSIM(db) + MSU_Blurring(db*) / 2
    * log(1-(encode_blur / source_blur))×(-10)

    if encode_blur < source_blur then (encode_blur / source_blur) - blurred image
    if encode_blur > source_blur then (source_blur / encode_blur) - sharp image

    And/or B-SSIM metric
    Last edited by zubzer0; 29th July 2016 at 06:40.

Page 2 of 2 FirstFirst 12

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •