Zero Frame Drop (was: Re: [XviD-devel] inter48x48)

Christoph Lampert xvid-devel@xvid.org
Sun, 28 Jul 2002 00:57:12 +0200 (CEST)


Hi,

finally, I have some results for a "Simpsons" Clip with many
"almost identical" frames (same image, just different noise). In total, it
was 16 frames out of 40, so 40% of all frames were "almost the same". 
The 16 is only the frames that are _copies of previous frames_, so the 
first frames of a sequence of identical images is _not_ counted in this
number, since it must not be removed. 

I encoded in normal mode (=NORM), with "identical" frames replaced by
exact copies (=ZERO) of the previous and completely without the "identical
frames" (=DROP). 

        size        gain              PSNR     speed      
Quant 2:

NORM: 352880                         44.60   123.0 fps  
ZERO: 290080     -17.7%              44.70   132.6 fps  +7.8%
DROP: 281280     -21.6%   -3.0%      44.59   113.5 fps  -7.7%


Quant 4:
        
NORM: 153480                         40.43   132.6 fps  
ZERO: 132480     -15.8%              40.43   136.5 fps  +2.9%
DROP: 128544     -18.8%   -2.9%      40.30   123.3 fps  -7.0%


Quant 8:

NORM:  62080                         36.38   140.8 fps
ZERO:  58520      -5.7%              36.34   143.2 fps  +1.7%
DROP:  55368     -10.8%   -5.3%      36.18   130.6 fps  -7.2%


Please note that DROP does _not_ include the bytes for marking the frame
as dropped, or the general AVI overhead in this case, but simply is the 
clip encoded if the frames in questions are completely deleted. So in
real life some bytes have to be added to the DROP values. 

Nevertheless there is a result and I guess everybody was somehow right: 

>From dropping 40%(!) "almost identical" frames you gain much less in
bitrate, between 10% and 20% depending on compression ratio (quant). 
The difference between encoding exact copies and dropping them is even
much lower: 3% to 5%. 

However, only for small qualitizers almost the same gain if achieved if
frames are set to exact copies (because internal image is good after first
compression step already and hardly any bits are needed in further
steps). For high quantizers, internal and external image differ too much
and encoding the exact copy still needs more bits. 

Encoding Quality is best if exact copies are encoded. No wonder, since the
encoder can spend extra bits in a second or third try to remove the
errors from first encoding step. 

There is one definite argument in favour of DROPping instead of 
only zeroing: Speed! Although relave speed drops (because only the
difficult frames have to be encoded, the most easy ones are left out), 
the absolute speed is of course much higher, if only 40% less frames have
to be encoded. 

Christoph 

-- 
Christoph H. Lampert chl@math.uni-bonn.de | Diese Signature wurde maschi-     
Beringstr. 6, Raum 14 Tel. (0228) 73-2948 | nell erstellt und bedarf
Sprechstunden: keine, aber meistens da    | keiner Unterschrift. AZ 27B-6