[x265] Slice decision (x264 vs x265)

Rainer M. Engel engel at endime.de
Wed Mar 1 16:40:40 CET 2017


Thank you. Then I understood correctly.

The reference has been created manually and secondly supported by a
cascaded feature based approach, where images next to slices would lack
of corresponding points.

Reference, 83593 frames, 848 scenes (incl. jump cuts etc.)
SCD (ref. above): 848 slices, 772 match with ref

Ranking so far:
01	x.264 	774 matching slices w. ref.
02	SCD	772 matching slices w. ref.
03	x.265	744 matching slices w. ref.  	

In a quality manner one can not say that the slice detection of x.265 is
poor. I was simply curious about the parameters and expected different
results by changing these parameters (see below). Maybe I will test that
again in the future.

Regards,
Rainer




Am 01.03.2017 um 11:26 schrieb Divya Manivannan:
> The expected behavior is the number of I-slices will increase with increase
> in scenecut-bias.
> 
> How are you identifying the reference (ground truth)?
> 
> Thanks,
> Divya M.
> 
> On Tue, Feb 28, 2017 at 8:33 PM, Rainer M. Engel <engel at endime.de> wrote:
> 
>> ... sorry for replying to my own mail..
>>
>> Test between x.264 and x.265 against reference (ground truth).
>>
>> Reference 848 scenes (slices)
>>
>> x.264.........................................................
>> x.264 | --min-keyint 1 --keyint 999999 --scenecut 50
>> found 839 slices, 766 match with reference
>>
>> x.264 | --min-keyint 1 --keyint 999999 --scenecut 100
>> found 910 slices, 774 match with reference
>>
>>
>> x.265.........................................................
>> x.265 | --keyint 999999 --scenecut 80
>> found 762 slices, 739 match with reference
>>
>> x.265 | --min-keyint 1 --keyint 999999 --scenecut 60 --scenecut-bias 15
>> found 799 slices, 744 match with reference
>>
>> x.265 | --min-keyint 1 --keyint 999999 --scenecut 100 --scenecut-bias 15
>> found 799 slices, 744 match with reference
>>
>> x.265 | --min-keyint 1 --keyint 999999 --scenecut 100 --scenecut-bias 30
>> found 795 slices, 744 match with reference
>>
>>
>> In these first tests I focussed on flexibility in setting the parameters
>> and aimed on more positive false detections if this increases the match
>> count against reference. According to that x264 produces higher rates.
>> Do --scenecut and --scenecut-bias already work like expected or am I
>> missing something here?
>>
>> Regards,
>> Rainer
>>
>>
>>
>> Am 25.02.2017 um 19:05 schrieb Rainer:
>>> Hello everyone,
>>>
>>> I'm new to this list and very interested in the "Slice decision options"
>>> of x265. The tests done so far show an expected behaviour of..
>>>
>>> --open-gop
>>> which is enabled by default.
>>>
>>> --keyint
>>> Setting this to a very high value, one which is higher than the duration
>>> of the longest shot, works fine.
>>>
>>> I run into problems by setting up the parameters to find all slices from
>>> my reference (ground truth containing any flavour of slice one can
>>> imagine:). Increasing "--scenecut" beyond the default of 40 brought no
>>> significant gain of more detected slices. Also increasing
>>> "--scenecut-bias" (as I understood correctly) wasn't of that big impact
>>> in the tests I've done so far.
>>>
>>> Are there some dependencies between parameters which are not that
>>> obvious here, at least to me? I also tried other parameters with no big
>>> impact.
>>> I hoped to get a feeling how the parameters work in concert, but the
>>> "aggressive" placement of slices by changing "--scenecut" values could
>>> not be reproduced here. Am I doing anything wrong?
>>>
>>> AIM:
>>> - individual slice detection by threshold, not interval
>>> - detect more false positives, than less
>>>
>>> CSV:
>>> - the csv log is used for comparing/counting the slices
>>>
>>> Any help, hint and reply on this is much appreciated.
>>> I use a MSVC build for Win-x64 found here http://x265.ru/en/builds/ on a
>>> YUV input clip.
>>>
>>> Best regards,
>>> Rainer
>>>
>>> --
>>> Rainer M. Engel, Dipl. Digital Artist
>>> Berlin, Germany


More information about the x265-devel mailing list