Effects of b-frames and qpel @ Q2

[Manao]

In reaction to what Tommy Carrot and Teegedeck posted in the DivX 5.11 vs XviD thread, I made a small test, @ Q2, with a b-frames settings 2,1,0 ( in order to have a quant 2 for b-frames ). All the other settings were set not to interfer ( no AQ, no QPel, no GMC, no Treillis, no Chroma Motion, no Chroma Optimizer, but VHQ4 ).

The movie was Hulk, I took a small sample of 100 frames in a low motion scene, I didn't resize nor filter ( just cropped ).

I encoded it with and without b-frames, and with and without QPel.

I obtained the following PSNR :

******* No BF, No QPel *******

         Absolute     Mean    Pos.   Neg.            
 Frame     Dev.       Dev.    Dev.   Dev.  PSNR (dB) 
-----------------------------------------------------
     0    0.4419    -0.1035    5     -5     50.8989
     1    0.6272    +0.0091    6     -7     48.6549
     2    0.6451    +0.0149    6     -7     48.5003
     3    0.6989    -0.0031    8     -8     47.8729
     4    0.6559    +0.0137    7     -8     48.4684
     5    0.6599    +0.0140    8     -7     48.4064
     6    0.7054    +0.0037    6     -7     47.8743
     7    0.6569    +0.0155    6     -7     48.4860
     8    0.6659    +0.0182    7     -7     48.3687
     9    0.7123    -0.0055    7     -7     47.8348
    10    0.6556    +0.0186    6     -7     48.4598
    11    0.6545    +0.0076    7     -8     48.4603
    12    0.6897    +0.0070    6     -7     47.9462
    13    0.6434    +0.0101    7     -8     48.5610

Mean PSNR : 48.2493

****** BF, No QPel *******

         Absolute     Mean    Pos.   Neg.            
 Frame     Dev.       Dev.    Dev.   Dev.  PSNR (dB) 
-----------------------------------------------------
     0    0.4419    -0.1035    5     -5     50.8989
     1    0.7005    +0.2978    9     -7     47.9595
     2    0.7019    +0.2943    7     -7     47.9433
     3    0.6584    +0.0006    8     -8     48.2278
     4    0.7188    +0.2716    7     -7     47.8259
     5    0.7243    +0.2896    9     -6     47.7732
     6    0.6793    +0.0197    7     -7     48.1729
     7    0.7342    +0.2868    7     -6     47.7368
     8    0.7304    +0.2529    8     -6     47.7897
     9    0.6768    +0.0094    6     -7     48.2496
    10    0.7157    +0.3253    8     -5     47.9418
    11    0.7212    +0.3248    8     -8     47.8843
    12    0.7111    +0.0173    7     -7     47.7750
    13    0.7177    +0.2623    7     -7     47.8605

Mean PSNR : 47.8908

******* No BF, QPel *******

         Absolute     Mean    Pos.   Neg.            
 Frame     Dev.       Dev.    Dev.   Dev.  PSNR (dB) 
-----------------------------------------------------
     0    0.4419    -0.1035    5     -5     50.8989
     1    0.6365    +0.0030    8     -8     48.3702
     2    0.6900    +0.0100    7     -6     47.8603
     3    0.7405    +0.0046    8     -8     47.3486
     4    0.7283    +0.0074    7     -8     47.5778
     5    0.7344    +0.0147    8     -7     47.5069
     6    0.7635    +0.0013    9     -9     47.1935
     7    0.7388    +0.0195    7     -7     47.5069
     8    0.7449    +0.0178    7     -7     47.4684
     9    0.7744    +0.0046    9     -7     47.1543
    10    0.7431    +0.0046    7     -7     47.4834
    11    0.7381    +0.0198    7     -7     47.5351
    12    0.7660    +0.0091    8     -8     47.1329
    13    0.7436    +0.0169    8     -7     47.4220

Mean PSNR : 47.3336

The results found were astonishing at first sight. The PSNR is lowered by BF and QPel, even at quant 2.

Let's begin by the b-frames :

The first explanation of this fact that come to mind ( distance between p-frames is raised, so motion compensation is worse, so PSNR will be worse for them, and hence PSNR for b-frames too ) is wrong.

In fact, as you can see, p-frames have a higher PSNR in the BF clip than if in the No BF one. But b-frames have a lower PSNR. I thought at first that I made a mistake, but after a second thought, I now think I understand : b-frames are built from the motion compensation of two frames, hence they are less sensitive to noise. And that is sufficient to make a difference : a drop by 0.3 ( from 48.2 to 47.9, mean PSNR )

The good news here is that the filesize is lower ( from 1260 KB to 1084 KB )

Now with QPel :

There again, PSNR is lowered, by quite a lot. There again, the explanation is the same : interpolation made by qpel works like a denoising filter. But the filesize is increased ( to 1550 KB ).

Conclusions : visually, there are no differences ( PSNR are too high for that ). The matrix used doesn't modify results ( I also tried H263, and an inter matrix filled with 8 ). And that's all. We can't deduce the behavior at higher quants, and QPel and BF are indeed useful when doing 'real' encodes.