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

[Marc Schulz]

Could you provide samples of your source?

I'm interested to see the quality difference between the two with such a 
drastic drop in the bitrate at the given crf value. I would imagine the 
quality drop to be substantial.


Vladimir Chernyshov wrote:
> Please don't misunderstand me, I am not asking for help. I am suggesting 
> a change in default options which can help an average user to get a 
> smaller bitrate without undertaking lengthy tests.
>
> Please allow me to disagree with your assertion of the encoder's 
> intelligence. The qcomp mechanism fails miserably in the scenes, where 
> part of the scene is highly active, and the other part (which can well 
> be the focus of viewer's attention) is essentially still. The qcomp will 
> consider the frame active as a whole and will kill the overall quality. 
> The qcomp thing also reduces sharpness in highly active scenes, and 
> although a viewer cannot distinguish the lost detail (the presumption of 
> the "constant quality" model), it can easily notice the loss of overall 
> sharpness and texture.
>
> This is why I used pure CQP mode for the video.
>
> However this is beyond the point of our discussion. Just for the sake of 
> the argument, I also repeated the test with and without scenecut 0 using 
> the default value of qcomp:
>
> --bframes 2 --b-adapt 0 --crf 18 --ref 3 --weightb \
> --keyint 250 --no-cabac --progress -o out1.264 -v raw-d2.1.y4m 
>
>
> x264 [info]: slice I:18    Avg QP:19.81  size: 49557  PSNR Mean Y:44.56 
> U:48.31 V:47.56 Avg:45.41 Global:45.31
> x264 [info]: slice P:227   Avg QP:21.93  size: 37747  PSNR Mean Y:43.12 
> U:47.24 V:46.33 Avg:44.03 Global:43.98
> x264 [info]: slice B:145   Avg QP:24.07  size: 29972  PSNR Mean Y:40.47 
> U:44.86 V:43.98 Avg:41.43 Global:41.37
> x264 [info]: consecutive B-frames: 36.8% 18.8% 44.4%
> x264 [info]: mb I  I16..4:  1.5%  0.0% 98.5%
> x264 [info]: mb P  I16..4:  2.1%  0.0% 80.6%  P16..4:  7.0%  8.3%  2.0% 
>   0.0%  0.0%    skip: 0.0%
> x264 [info]: mb B  I16..4: 22.6%  0.0%  0.0%  B16..8: 28.5% 24.2% 18.7% 
>   direct: 6.0%  skip: 0.0%  L0:46.3% L1:38.4% BI:15.4%
> x264 [info]: ref P L0  71.0% 17.2% 11.8%
> x264 [info]: ref B L0  72.9% 27.1%
> x264 [info]: SSIM Mean Y:0.9683855
> x264 [info]: PSNR Mean Y:42.201 U:46.405 V:45.516 Avg:43.127 
> Global:42.859 kb/s:14160.45
>
> -----------------------------------------------------------------------
> --bframes 2 --b-adapt 0 --crf 18 --scenecut 0 --ref 3 --weightb \ 
>
> --keyint 250 --no-cabac --progress -o out1.264 -v raw-d2.1.y4m 
>
>
> x264 [info]: slice I:2     Avg QP:17.61  size: 62501  PSNR Mean Y:46.19 
> U:49.50 V:48.89 Avg:46.98 Global:46.63
> x264 [info]: slice P:130   Avg QP:21.90  size: 38137  PSNR Mean Y:43.24 
> U:47.37 V:46.48 Avg:44.15 Global:44.10
> x264 [info]: slice B:258   Avg QP:24.18  size: 30687  PSNR Mean Y:40.36 
> U:44.75 V:43.91 Avg:41.32 Global:41.25
> x264 [info]: consecutive B-frames:  0.3%  0.0% 99.7%
> x264 [info]: mb I  I16..4:  1.1%  0.0% 98.9%
> x264 [info]: mb P  I16..4:  2.4%  0.0% 87.2%  P16..4:  4.5%  5.0%  0.9% 
>   0.0%  0.0%    skip: 0.0%
> x264 [info]: mb B  I16..4: 23.7%  0.0%  0.0%  B16..8: 27.5% 24.9% 18.7% 
>   direct: 5.3%  skip: 0.0%  L0:48.4% L1:38.3% BI:13.3%
> x264 [info]: ref P L0  57.7% 22.3% 20.1%
> x264 [info]: ref B L0  70.4% 29.6%
> x264 [info]: SSIM Mean Y:0.9627631
> x264 [info]: PSNR Mean Y:41.349 U:45.646 V:44.793 Avg:42.294 
> Global:42.027 kb/s:13333.52
>
> encoded 390 frames, 3.70 fps, 13334.15 kb/s
>
> As you can see, the scenecut=0 mode still gives the best bitrate by 
> properly utilizing B-frames.
>
> regards,
> Vladimir
>
>
> Jason Garrett-Glaser wrote:
>   
>>> Thus with plenty of I- and P- frames the average quality is boosted,
>>> which is against the idea of CQP VBR mode, and not desirable in
>>> highly-active scenes.
>>>       
>> The entire point of constant-quality VBR mode is to reduce bitrate
>> spent in complex scenes... which you have proceeded to disable by
>> setting qcomp to 1.
>>
>> I am not going to bother helping people who explicitly override
>> intelligent encoder decisions with misinformed ones of their own and
>> then complain when they get suboptimal results.
>>
>> Dark Shikari
>> _______________________________________________
>> x264-devel mailing list
>> x264-devel at videolan.org
>> http://mailman.videolan.org/listinfo/x264-devel
>>
>>     
>
> _______________________________________________
> x264-devel mailing list
> x264-devel at videolan.org
> http://mailman.videolan.org/listinfo/x264-devel
>
>