[x264-devel] Default value for the --scenecut option causes heavy undesirable flood of I slices in active scenes

Vladimir Chernyshov vchernys at welho.com
Tue Nov 4 10:24:53 CET 2008


I don't have a better idea how to detect highly active scenes where  
bits can be saved, as opposed to those where saving is harmful. Like  
many tasks from the domain of the computer vision, this may be  
impossible to achieve anyway.

However, the proportion of highly active scenes in a typical video is  
usually not big, and in VBR mode it is perfectly OK to allocate as  
many bits to these scenes as needed. So the crf=18 qcomp=1.0 mode with  
B-frames worked well for me.

The only problem was that the bitrate was overinflated by the flood of  
I-frames in highly active scenes. E.g. 22 Mbit/s was inflated to 30  
Mbit/s (the original was a 25 Mbit/s interlaced DV video, and the  
average bitrate of the resulting video was supposed to be 12 Mbit/s).  
This spike makes it hard to predict CPU and transport requirements for  
playing the video.

I had to disable CABAC because my computer would not have the power to  
play the video encoded with anything close to the desired quality.

I did not disable AQ, I value how much it improves video quality at  
the given crf.

I repeat the point of my posts: it is counterproductive to insert  
I-frames and suppress B-frames in highly active scenes - the bitrate  
would be inflated, and the quality improvement not noticed, and the  
default value for the scenecut could be changed from 40 to 0.

regards,
Vladimir


>> Patches to implement my "MB-tree" algorithm to solve this problem (and
>> make qcomp no longer necessary) are welcome.
>>
> Also note the following:
>
> 1.  Let us say that qcomp raises QP by 2 in that section, and forcing
> B-frames raises QP by 2 as well.  The latter will always be less
> efficient, because more B-frames will make that scene less efficiently
> coded--guaranteed.  If what you want is a hierarchical QP structure,
> again, you can get that without B-frames, and it will be more
> efficient as well.
>
> 2.  Forcing QP mode disables AQ, which will cause simple parts of the
> frame to be trashed even further in favor of the more complex areas of
> the frame.
>
> 3.  I know of no H.264 encoder in the entire world that implements MB Tree.
>
> Dark Shikari




More information about the x264-devel mailing list