[FFmpeg-devel] [PATCH] avfilter: Added siti filter

[Thilo Borgmann]

Am 18.01.22 um 14:58 schrieb Thilo Borgmann:
> Am 15.01.22 um 09:27 schrieb Paul B Mahol:
>> On Sat, Jan 15, 2022 at 9:25 AM Paul B Mahol <onemda at gmail.com> wrote:
>>
>>>
>>>
>>> On Fri, Jan 14, 2022 at 5:22 PM Thilo Borgmann <thilo.borgmann at mail.de>
>>> wrote:
>>>
>>>> Hi,
>>>>
>>>> Am 19.01.21 um 05:49 schrieb Lynne:
>>>>> Jan 19, 2021, 01:07 by borbarak at fb.com:
>>>>>
>>>>>> Calculate Spatial Info (SI) and Temporal Info (TI) scores for a video,
>>>> as defined
>>>>>> in ITU-T P.910: Subjective video quality assessment methods for
>>>> multimedia
>>>>>> applications.
>>>>>>
>>>>>> Update: Fixed bracket style.
>>>>>>
>>>>>
>>>>> Thanks, looks much neater now.
>>>>>
>>>>>
>>>>>
>>>>>> I'm already adding the data to the frame's metadata, is the suggestion
>>>> to remove the file option altogether?
>>>>>>
>>>>>
>>>>> Yes. We want to avoid filters having their own file in/out options
>>>> rather
>>>>> than using generic ones.
>>>>
>>>> Updated the patch to apply to git HEAD.
>>>> Removed file output.
>>>> Made printing summary to console optional.
>>>>
>>>>
>>>>>> +
>>>>>> +#include "libavutil/imgutils.h"
>>>>>> +#include "libavutil/internal.h"
>>>>>> +#include "libavutil/opt.h"
>>>>>> +
>>>>>> +#include "avfilter.h"
>>>>>> +#include "formats.h"
>>>>>> +#include "internal.h"
>>>>>> +#include "video.h"
>>>>>> +
>>>>>> +static const int X_FILTER[9] = {
>>>>>> +    1, 0, -1,
>>>>>> +    2, 0, -2,
>>>>>> +    1, 0, -1
>>>>>> +};
>>>>>> +
>>>>>> +static const int Y_FILTER[9] = {
>>>>>> +    1, 2, 1,
>>>>>> +    0, 0, 0,
>>>>>> +    -1, -2, -1
>>>>>> +};
>>>>>>
>>>>>
>>>>> We have optimized assembly to apply 3x3 matrices. Check out
>>>>> libavfilter/x86/vf_convolution.asm:ff_filter_3x3_sse4
>>>>>   vf_convolution already applies a sobel filter that way. Maybe
>>>>> look into sharing some DSP code with it?
>>>>
>>>> I checked a bit since I also want a common sobel for vf_edgedetect / my
>>>> patch for vf_blurriness.
>>>>
>>>> For sobel, there is no direct asm implementation. We have a generic
>>>> filter_3x3 with sse4 optimization. To use sobel with that, you'd need to
>>>> run two times filter_3x3 plus one pass for gradient calculation.
>>>>
>>>> As another difference, this filter (SITI) does on-the-fly conversion to
>>>> full-range pixel values during its sobel. While vf_edgedetect /
>>>> vf_bluriness use an abbreviation for the gradients during its sobel. Which
>>>> makes them both distinct enough not to fit into a general filter_3x3 or
>>>> filter_sobel from vf_convolution easily (and more overhead). So I think
>>>> it's not worth the effort to force them into a common function? (Especially
>>>> since we don't have a sobel_sse4 or something)
>>>>
>>>> Patch without a common sobel attached.
>>>>
>>>>
>>>
>>> Violations of code style.
> 
> Enhanced.
> 
> 
>>> Also why filter description only shows SI part and TI part is missing?
> 
> Mentioned.
> 
> v2 attached.

Ping.

-Thilo