[XviD-devel] Mode decision

[Christoph Lampert]

Hi Skal,

hm, from a quick glance at your data I'd say: 

Sometimes trying as much INTER4V as possible is good (though slow),
sometimes trying any INTER4V is bad (and slow), sometimes it doesn't
matter at all (except for being slow). 

So be honest, this doesn't sound like a "we should take 60% as a
compromise" (even though this would mean only 2% too large file for
squares and less than 1% too large for Hollow Men), but it rather sounds
like "we should find out _when_ inter4v is beneficial and when it's not
and use that as a trigger". After all, there is not reason why INTER4V
should make things _worse_ at all in any case, except for a bad
implementation.  
If it cannot be detected automatically, we could at least add a switch
(anime<->reallife) and adapt the quality levels accordingly. 

Btw. can anyone confirm that DivX5.0.x _never_ uses INTER4V? 
I was astonished to hear that... 

gruel 


On 25 Mar 2003, skal wrote:
> 	Gruel,
> 
> On Mon, 2003-03-24 at 19:55, Christoph Lampert wrote:
> > On 24 Mar 2003, skal wrote:
> > > > What you can't see, is how much a misclassification of INTER->INTER4V
> > > > or vice versa would cost in extra bits. 
> > > > And you also cannot see that there are hardly any real INTER4V blocks
> > > > with sad16 below 600. So maybe this would be a way to speed things up, 
> > > > simply not doing inter4v for those.
> > > 
> > > 	Even with a rather crude sad16-vs-sad8 final criterion for
> > > 	INTER/INTER4v decision, I've tried to guess whether a 16x16
> > > 	INTER block was worth a candidate for a sub 4V search (which
> > > 	takes times). I've tried some criterion based of local gradient
> > > 	and divergence (is the MV-field torn appart?), but it's not
> > > 	convincing. Actually, it seems that 4V works best at image's
> > > 	segmentation limits. So far, so good, I ended up with a sad
> > > 	-based criterion that is a good hint of whether going for a
> > > 	sub search might be rewarding: after a 16x16 regular search,
> > > 	I re-use the best MV found so far to evaluate sad8 of each of
> > > 	the four 8x8 sub blocks. If the maximum of these sad8 values 	is
> > > greater than a fraction the sad16 for the full block
> > > 	(in practice 60% is a good compromise), then I go for a
> > > 	refined search...
> > > 
> > > 	Any opinion?
> > 
> > Yes, it might be possible to check earlier if INTER4V is promising
> > (INTER is already very good => no, SAD8 very different => yes).
> > I had plans using that current search also find best positions and
> > SADs so far for the 8x8 blocks during search for 16x16.  But my results
> > were not convincing.
> > 
> > If yours is better, great! Do you have results? 
> 
> 	Here are some results for two sequences I think are
> 	typical: first one (squares) is made of moving squares
> 	overlapping each other (might resemble an "anime"),
> 	and the other is made of 400 frames taken from the
> 	"Hollow man" trailer. The squares sequence was generated
> 	using the attached small program (use: warp_gen -t 2 -n 400).
> 
> #%probing|file size |FPS
> 
> # moving squares (~anime?), 400 frames
> 0	  5863498    240.964
> 10	  5809766    235.294
> 20	  5691258    232.558
> 30	  5556203    224.719
> 40	  5376767    218.579
> 50	  5118033    209.424
> 60	  4884111    199.005
> 70	  4797765    188.679
> 80	  4796947    186.916
> 90	  4796455    184.332
> 100	  4794623    186.047
> 
> 
> # hollow-man 640x352, 400 frames w/ scene changes
> 0	  3708908    86.022
> 10	  3708911    85.837
> 20	  3709495    85.470
> 30	  3709517    85.653
> 40	  3711307    85.288
> 50	  3715019    84.926
> 60	  3738086    81.466
> 70	  3840476    70.796
> 80	  3826657    65.789
> 90	  3826657    66.007
> 100	  3826657    66.007
> 
> 	Here are also the figures for the "parkrun" sequence:
> # parkrun 1280x720, 33 frames
> 0 	10332336 10.476
> 10	10332244 10.377
> 20	10332354 10.061
> 30	10331736 10.217
> 40 	10324032 10.185
> 50	10319714 10.000
> 60      10308481 10.123
> 70	10293848 8.824 
> 80	10307127 7.820 
> 90	10307127 7.838 
> 100     10307127 7.857 
> 
> 
> 	x="% probing" is the threshold used to decide
> 	going into the 4v search: if max{sad8}>(1-x)*sad16
> 	=> try 4v mode.
> 
> 	The curves exhibit two behaviors: "square" sequence
> 	is a very good candidate for 4v, but one have to
> 	probe enough blocks (50-60%). However, too much
> 	probing is time consuming for few improvements.
> 	On the contrary, the Hollow-man sequence is a very
> 	bad candidate for 4v, no matter what. In fact,
> 	file size with full 4v-testing is *bigger*, because
> 	of the crudeness of the sad16-vs-sad8 criterion you
> 	were talking about.
> 
> 	=> I think the good compromise is around 50-60% probing.
> 
> 	admittedly, more test sequences is needed...
> 
> 	bye,
> 			Skal
> 
> (PS: the same kind of results seems to apply to interlaced field prediction...)
>