On Jun 3, 2014, at 12:18 PM, Laurent Birtz <[email protected]> wrote:

> 
>> 1) I don't agree with the fact that small blocks matter more.
>> The ETS lab ran tests with the HM and demonstrated that as the QP increases, 
>> smaller blocks are neglected at the profit of bigger ones. At QP 37, more 
>> than 50% of the blocks are 64x64. This is true for the HEVC test sequences 
>> of classes B (1920x1080), C (832x480) and E (1280x720). This ratio is less 
>> for the sequences in class D (416x240) where no more than 40% of the blocks 
>> are 64x64.
> 
> The blocks that did get split may have consumed the majority of the bits, 
> though. Not splitting them may cause the residual to grow very large. We'll 
> see in practice.

The tests simply observed the percentage of CTBs of each size. They didn’t 
observed the differences between the cost of using a 2Nx2N CU against four NxN 
CUs. I do agree with your logic that not merging blocks that should have been 
merge will have less of an impact than not splitting a CU that should clearly 
have been. The latter case would result in either a steep increase in bit-rate, 
or very poor quality because of aggressive quantization. We want neither of 
these options.


> 
> 
>> 2) More experiments need to be carried out to make sure replacing context 
>> bins by a fixed value of 50% isn't a false good idea. Observations were 
>> carried out at QP 30. Do they hold at QP 20 and QP 40?
> 
> Yes. But I expect that if it works for QP 30, then it's going to keep working 
> at least for one QP direction (up or down, I don't know which).
> 

Just for the sake of science, we should test it before. A couple of videos 3 ~ 
4 should be enough.

> 
>> 3) I would probably also look at the intra 16x16 modes selected prior to 
>> doing any processing at the 32x32 level. We should look for a common 
>> direction/mode. My gut feeling tells me that if all modes are very 
>> different, intra 32x32 will rarely win. I think that this will be especially 
>> useful to skip intra 32x32 at borders where intra 16x16 are better suited to 
>> capture the signal characteristics.
> 
> Yes, that's the correlation test I was referring to.
> 
> 
>> 4) Did you mean 32x16 and 16x32 searches in "then the 32x32
>> search is performed, and possibly 16x8 and 8x16 too, following the same logic
>> as the 16x16 CB.” ? AFAIK, inter 16x8 and 8x16 would have been carried out 
>> by the time we get to the 32x32 level. In fact, AMP look to be triggered 
>> when 16x8 or 8x16 are competitive.
> 
> Yes, typo.
> 
> 
> Thanks for the review.
> Laurent

NP
François

> --
> To unsubscribe visit http://f265.org
> or send a mail to [email protected].
> 

--
To unsubscribe visit http://f265.org
or send a mail to [email protected].

Reply via email to