[x264-devel] Thresholds about slicetype_decide

[kenter]

I find there is a bug in my hack, after bug fixing, the PSNR is
34.877dB(vs 34.855dB of 8x8 ME version)

However, I still have some questions:
1. There is only 0.022dB PSNR gain for slicetype_decide by 16x16 vs
8x8 lowres one, Is it reasonable?
2. The return value of x264_slicetype_mb_cost must be divided by 4
(i_bcost/4), otherwise, there is a 0.5dB PSNR drop, I think it doesn't
make sense only if there are some parameters/thresholds related to
8x8, can you help me to find out these parameters/thresholds?

Thanks!

2010/5/19, kenter <kenter83 at gmail.com>:
> The process of my hack as below:
>     1. Interp. half-pel values of origin image(at the same time of
> x264_frame_init_lowres)
>     2. Analyse each MB(Set QP=24) and sum-up cost(16x16ME instead of
> 8x8; Intra analyse also modified as Intra16x16 prediction), the
> returned value is divided by 4.
>     3. slicetype_decision & cost comparison between references are the
> same as before.
>
> The data array to store lowres costs also modified as uint32_t to
> assure it'll NOT overflow.
>
> 2010/5/19, kenter <kenter83 at gmail.com>:
>> yes, the Intra prediction have been modified as 16x16 in
>> slicetype_decide, and the returned i_cost of slicetype_decide is also
>> divided by 4(I assume that SATD(16) = 4*SATD(8))
>>
>> 2010/5/19, Jason Garrett-Glaser <darkshikari at gmail.com>:
>>> On Tue, May 18, 2010 at 10:05 PM, kenter <kenter83 at gmail.com> wrote:
>>>> Thanks for Jason, I'm sorry that I forget to mention my modifications
>>>> are based on x264-snapshot-20090302-2245 for historical reasons, the
>>>> mb-tree and lookahead feature are not available.
>>>> I modulate the parameters based on your suggestions:
>>>> 1. Set QP=24 in x264_lowres_context_init()
>>>> 2. Change the data array type of *i_intra_cost from uint16_t to
>>>> uint32_t in x264_frame_t (It seems the lowres_cost and lowres_type are
>>>> not exist in my version)
>>>>
>>>> However, the result shows the PSNR also drops 0.5dB compares to that
>>>> of 8x8 slicetype_decide(from 34.84 to 34.35dB), can you point me other
>>>> values that I need to modify?
>>>
>>> You'll need to redesign intra analysis in lookahead if you want to use
>>> 16x16 blocks; currently it assumes 8x8 and uses 8x8 intra prediction
>>> accordingly.
>>>
>>> Dark Shikari
>>> _______________________________________________
>>> x264-devel mailing list
>>> x264-devel at videolan.org
>>> http://mailman.videolan.org/listinfo/x264-devel
>>>
>>
>