[FFmpeg-user] Cut detection with dynamic threshold possible?

[Paul B Mahol]

On 4/26/16, Christoph Gerstbauer <christophgerstbauer at gmail.com> wrote:
>
>
>
>>> On Apr 25, 2016, at 11:25 AM, Christoph Gerstbauer
>>> <christophgerstbauer at gmail.com> wrote:
>>>
>>> Hello,
>>>
>>> I want to make shotdetections (scene cut) with ffmpeg. (a shot is the
>>> video between 2 cuts)
>>>
>>> My syntax knowledge at this time offers to generate only the FIRST FRAME
>>> of each shot. NO LAST FRAME. It would be nice to find also the last frame
>>> of a shot, but anyway this ticket should focus on the threshold itself.
>>>
>>> Actual Syntax to generate shotdetection thumbnails with ffmpeg:
>>>
>>> ffmpeg -i <inputvideo> -vf select='gt(scene\,0.7)' -vsync 0 -an
>>> keyframes%03d.jpg
>>>
>>> the value "0.7" is the threshold.
>>> A higher one will lead to less scenes thumbails - low sensitivity
>>> A lower one will lead to much more thumbnails - high senisitivity
>>>
>>> I mentioned that shotdetection algorythms generally has problems with to
>>> dark scenes, so it would be fine to have a threshold which is DYNAMIC to
>>> find every cut in the video. But my "FFmpeg-Fu" is to weak to find a
>>> solution on my own.
>>> So, is it possible to use a dynamic threshold within ffmpeg scene
>>> detection, and  if yes -> how can I do that?
>> You could also consider using the YDIF values from the signalstats filter.
>> It will quantify the visual difference in the Y plane from one frame to
>> the next and thus spike on scene cuts.
>> Dave Rice
>
> Hell Dave,
> thank you for the hint. I am testing now different videos with QC tools
> (based on ffmpeg) and using the integrated Y filter to show a graph on
> the timeline. That works good. Looks like in Davinci Resolve scene cut
> filter :) And the peaks in the timeline hit hardly all hard cuts. Some
> false positives when flashing happens in video, but this does not matter.
>
> But what I need is a syntax which I can use in FFmpeg commandline which
> uses the ydif function from the singalstats filter and force ffmpeg to
> output jpgs at the video position where this big peaks happen. But as I
> said, I have no idea how I can do it :/
> I need a working formular/syntax for that. Do you have any idea how I
> this yntax would look like?

ffmpeg -i INPUT -vf
signalstats,metadata=mode=select:key=lavfi.signalstats.YDIF:value=25:function=greater
OUTPUT