[FFmpeg-devel] [FFmpeg-cvslog] Adds ESPCN super resolution filter merged with SRCNN filter.

Pedro Arthur bygrandao at gmail.com
Wed Jul 4 17:19:22 EEST 2018


2018-07-04 4:03 GMT-03:00 Jean-Baptiste Kempf <jb at videolan.org>:
> On Wed, 4 Jul 2018, at 01:22, Carl Eugen Hoyos wrote:
>> 2018-07-04 0:14 GMT+02:00, Jean-Baptiste Kempf <jb at videolan.org>:
>> > On Tue, 3 Jul 2018, at 20:59, Carl Eugen Hoyos wrote:
>> >> How is this case different from many arrays in libavcodec/*data*?
>> >
>> > It is very different: the arrays in *data* come either from a
>> > mathematical computation or a spec.
>>
>> (Apart from: Free and open specs?)
>> This is probably true for some of the arrays, I think it is
>> very unlikely that it's true for all of them.
>
> The point is: you can recreate all of those arrays.
> If OP dies, you can still take over.
>
>> > Else, as some Debian Developer said: "It looks like code
>> > hidden in an unsigned char array"
>>
>> Is it "code" or data that was computed with a copyrighted
>> algorithm?
It is only data, namely the weights used in the filters.

> How can you know, if it is not explained, and you cannot reproduce it?
> How is it different from a binary blob?
>
> Anything related to NN is very annoying, if you don't share the dataset and the methodology.

The paper cited in the code contains all relevant information for
anyone with basic CNN knowledge to understand the model and maintain
the code.
The whole point of training a NN is that you do it once, takes a lot
of time, and never do it again.
If you look at the publications in this area, the concept of
reproducibility is not the usually expected "bit exact" ,as other
mathematical fields.
It is more like achieving similar results, given the model and dataset.

For anyone interested in the training process, Sergey provided the
repo [1] with scripts for downloading the dataset and training.

[1] - https://github.com/HighVoltageRocknRoll/sr


More information about the ffmpeg-devel mailing list