[FFmpeg-user] datascope 16-bit

[Jim DeLaHunt]

On 2022-11-20 00:15, Michael Koch wrote:

> Am 20.11.2022 um 02:39 schrieb list+ffmpeg-user at jdlh.com:
>>
>> On 2022-11-19 07:34, Michael Koch wrote:
>>> Am 03.11.2022 um 10:10 schrieb Paul B Mahol:
>>>>
>>>> Some things in sea of myriad others:
>>>> [...]
>>>> Why you claim that datascope does not support >8 bit formats where in
>>>> fact it does support it?
>>>
>>> In some cases datascope works with 16-bit data, but not in all cases.
>>> Here is an example where it doesn't work:
>>>
>>> ffmpeg -f lavfi -i color=black:s=32x32,format=gray10le -lavfi 
>>> geq=lum='X+32*Y',format=gray10le -frames 1 -y test.png
>>>
>>> ffmpeg -i test.png -vf 
>>> format=rgb48,showinfo,datascope=s=1280x1152:mode=color2:format=hex 
>>> -y out.png
>>>
>> Perhaps clarify what you observe as "doesn't work", and what 
>> behaviour you expect?
>>
>> Both those commands run for me without error, and I can view both 
>> test.png and out.png without problem. out.png has a cascade of 32 
>> 2-digit hex numbers on the left half, and solid black on the right 
>> half. The hex numbers run from 00 to 7F, in white on a black to grey 
>> gradient background, and from 80 to FF, in black on a grey to white 
>> background.
>
> I would expect 4-digit hex numbers, because the rgb48 pixel format is 
> 16-bit per channel.
> For example, it works fine if "rgb48" is replaced by "gray16".

The phrase "works fine" appeals to some notion of what is "correct" 
behaviour. It seems that you have your own idea of "correct" behaviour 
for this filter. But it seems more helpful for communication on this 
list to use FFmpeg's idea of "correct" behaviour. And the best source we 
have for FFmpeg's idea of "correct" is the filter documentation 
<https://ffmpeg.org/ffmpeg-all.html#datascope>: "Video data analysis 
filter. This filter shows hexadecimal pixel values of part of video."

I think the description in the documentation is incomplete and unclear. 
I wish FFmpeg had a better description. But the actual behaviour does 
not conflict with this incomplete description. The description does not 
promise that the datascope filter shows the full-precision, untruncated 
pixel values. It might be (I did not check) that the 8-bit values which 
datascope displays for an rgb48 input image are the correct upper 8 bits 
of 16-bit pixel values.

So, you said, "In some cases datascope works with 16-bit data, but not 
in all cases."  If you had instead said, "In some cases datascope gives 
useful results with 16-bit data, but not in all cases", then I would be 
completely with you. It is clear that truncated 8 bit values for an 
rbg48 input are not as helpful as full-precision, untruncated 16-bit 
pixel values.

But the sad reality is the FFmpeg only does not always document its 
intended behaviour clearly, and does not seem to have a goal of always 
providing the most helpful behaviour. The culture here understands 
"doesn't support" or "doesn't work" to mean, "ffmpeg terminates 
prematurely with an error" or "ffmpeg fails to generate output". If you 
use those phrases when your objection is actually, "runs to completion 
and generates output, but output is not as helpful as it could be", then 
your message gets diluted by misunderstanding.

Best regards,
      —Jim DeLaHunt