[FFmpeg-user] Why is format=rgb24 required after maskedmerge?

[Michael Koch]

Am 19.08.2020 um 16:34 schrieb Paul B Mahol:
> On 8/19/20, Michael Koch <astroelectronic at t-online.de> wrote:
>> Am 18.08.2020 um 14:35 schrieb Michael Koch:
>>> Hello all,
>>>
>>> I have a question about this script:
>>>
>>> rem  Create red and yellow videos 300x300
>>>
>>> c:\ffmpeg\ffmpeg -f lavfi -i color=red:size=300x300:duration=5 -y red.mp4
>>> c:\ffmpeg\ffmpeg -f lavfi -i color=yellow:size=300x300:duration=5 -y
>>> yellow.mp4
>>>
>>> rem  Create a mergemap file 600x300
>>>
>>> c:\ffmpeg\ffmpeg -f lavfi -i nullsrc=size=300x300 -vf
>>> "format=gray8,geq='clip(128-128/10*(180-200/150*hypot(X-150,Y-150)),0,255)',v360=input=fisheye:output=e:ih_fov=200:iv_fov=200,format=rgb24"
>>>
>>> -frames 1 -y mergemap.png
>>>
>>> rem  Stitch two fisheye videos together to an equirectangular video,
>>> with merging at the border
>>>
>>> c:\ffmpeg\ffmpeg -i red.mp4 -i yellow.mp4 -i mergemap.png -lavfi
>>> "[0]format=rgb24,v360=input=fisheye:output=e:ih_fov=200:iv_fov=200[a];[1]format=rgb24,v360=input=fisheye:output=e:yaw=180:ih_fov=200:iv_fov=200[b];[a][b][2]maskedmerge,format=rgb24,format=yuv422p"
>>>
>>> -y out.mp4
>>>
>>>
>>> This works fine and the output video looks as expected. My question is
>>> why "format=rgb24" is required between "maskedmerge" and
>>> "format=yuv422p". Normally two consecutive format conversions should
>>> be unnecessary. But it doesn't work when I omit "format=rgb24". The
>>> colors become wrong.
>> It seems that v360 and maskedmerge change the pixel format from rgb24 to
>> the planar gbrp pixel format.
>> But why isn't it possible to convert gbrp directly to yuv422p, without
>> an intermediate conversion to rgb24? Is this a bug?
>>
>> I suggest to add to the documentation that v360 and maskedmerge (and
>> possibly also other filters) change the pixel format to gbrp. As a user
>> I was assuming that the pixel format remains the same, unless noted
>> otherwise in the documentation.
> You are deeply confused about our filters.
> Any filter can change pixel formats to one that they accepts thus gbrp
> is picked instead of packed rgb, this is already documented
> implicitly.

It would be very helpful to find in the documentation for each filter a 
list of the accepted input pixel formats, and it's also a useful 
information if a filter is capable of changing the pixel format. Or can 
all filters change the pixel format?

> Your mergemap is produced in rgb24 pixel format and it only work with
> rgb pixel formats otherwise you get wrong output from maskedmerge.

OK, understood. I converted the mergemap to gbrp and now it works fine.
In my opinion the whole thing is badly documented.
On one hand, the filter does automatically change the first two inputs 
to a format that it accepts, in this case gbrp.
On the other hand, the filter is unable to automatically change the 
third input to the same pixel format, which is the only pixel format 
that works.

Do you agree that this is the correct workflow?
1. Insert a showinfo filter after maskedmerge and check the pixel format.
2. Insert a format conversion in the filter chain that converts the 
mergemap to the same pixel format as found in step 1.

Michael