[FFmpeg-devel] [PATCH 2/2] libavfilter: Removes stored DNN models. Adds support for native backend model file format in tf backend.

[Pedro Arthur]

2018-08-21 16:40 GMT-03:00 Gyan Doshi <gyandoshi at gmail.com>:
> On 21-08-2018 10:27 PM, Pedro Arthur wrote:
>
>>>
>>>
>>> Something like,
>>>
>>>      ffmpeg -i in.bmp -filter_complex
>>>             'format=yuvj444p,extractplanes=y+u+v[y][u][v];
>>>              [y]sr=dnn_backend=tensorflow:model=espcn.model[y-sr];
>>>              [u]scale=iw*2:-1[u2x];[v]scale=iw*2:-1[v2x];
>>>              [y-sr][u2x][v2x]mergeplanes=0x001020:yuvj444p'
>>>              out.bmp
>>>
>> I personally think the average user, just wanting to upscale and
>> image, shouldn't have to type that much.
>
>
> Agreed. Can't the filter upscale the chroma planes internally?
That's exactly what the filter were doing before, but for some reason
doing it this way is not 'good code'.

I think having the power to define graphs of filters were each filter
does a single job is nice and elegant but in this case it is doing a
disservice to the user, it shouldn't be a strict rule when usability
is compromised.

>
>
> Gyan
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel