[Mp4-tech] [H.264] FrameSkip affect lambda_scale after JM9.3

[alexismt adelphia.net]

 Dear Sothat,
I would actually encourage you to test this concept with a bigger set of sequences but also considerations of hierarchical coding structures, and qp variations between different slice types or hierarchy. This concept has been tested before (especially in the context of hierarchical structures) and although it was found to be hurting slightly for some content, it was found helping considerably for several others. Since the RDO also in the software is also a "suggested" method and not an optimal method, therefore this concept was added in the software as part of JVT-N014.
In general this "trick" was added as a way to consider the temporal dependency and distances during MC. As you would have expected, increasing temporal distant would have an impact on ME and MC, which would finally possibly result in a worse reference for future pictures. As you can understand this would basically result into a chain reaction in terms of MC. Reducing the lambda based on temporal distances was a simple way to compensate for this behavior, improve references, but also improve overall performance.
I am not certain what conditions you are using, but it is quite common, especially in traditional structures of the form IBBPBBP... to use a larger QP for B slices than for P. This concept was added with this in mind as well (just in case if your QPs are set equal). 
Best regards,
Alexis
---- "Martin Y. Chiu" <sothat bsdx.org> wrote: 
> Dear all,
> 
>         We had some experiment about B-pictures on JM. However,
> 
>     there were strange phenomena when we have changed FrameSkip
> 
>     parameter in config file. It seems that was introduced after 
> 
>     JM 9.3 by the code @ rdopt.c:3002
> 
>     double lambda_scale = (img->type == B_SLICE &&
>         img->nal_reference_idc == 0 ? 1.0: 1.0 - 0.05 * (double)
>         input->jumpd);
> 
>     Also it is the same in JM 9.8 slice.c:1062
> 
>     double lambda_scale = 1.0 - Clip3(0.0,0.5,0.05 * (double)
>             input->jumpd);;
> 
>         What does it mean of the lambda_scale factor ?  In our 
>         
>     experiment, it introduced PSNR loss with this factor. Is there any
> 
>     docuemnt related to this parameter ? Thanks.
> 
> -- 
> Regards,
>     sothat
> _______________________________________________
> NOTE: Please use clear subject lines for your posts. Include [audio, [video], [systems], [general] or another apppropriate identifier to indicate the type of question you have.
> 
> Note: Conduct on the mailing list is subject to the Antitrust guidelines found at http://www.mpegif.org/public/documents/vault/mp-out-30042-Antitrust.php