Encoding Discussion Megathread (ffmpeg, Handbrake, AV1, etc...)

[TheDreadPirate]

Circling back to the cropdetect.  Two things

1)  I got accelerated decoding working with crop detect.
2)  I've recently re-encountered an edge case that I've encountered a few times in the past.

The accelerated command looks like this.  It doesn't make crop detection happen any faster, in fact it slows things down by about 50%, but it massively lowers the CPU usage.

Code:

ffmpeg -init_hw_device vaapi=va:,driver=iHD,kernel_driver=i915 -init_hw_device qsv=qs@va -filter_hw_device qs -hwaccel qsv -hwaccel_output_format qsv -ss 120 -t 5:00 -i Bookworm.mkv -vf hwdownload,format=p010le,fps=1/2,cropdetect=mode=black:reset=60 -f null - 2>&1

It's probably the hwdownload slowing things down.  But, as you mentioned, I couldn't find a QSV filter to do everything in hardware.

You will have to do some detection first with ffprobe.  The pixel format filter needs to match the video.  "p010le" for 10-bit video and "nv12" for 8-bit video are the two that QSV supports.

Regarding #2.  If, at the start of the cropdetect at the -ss timestamp, your video has a different aspect ratio, either for stylistic effect or because it is a small logo on an all black background, the cropdetect will get it wrong if the windows isn't large enough.  Your -t 5:00 helped me with that.  I had been using "-ss 120 -to 180".  MOST videos will have go back to 16:9 or 21:9 by the end of that 5 minute window.

But I have a new movie where there is an extended intro/prelude with a different resolution and aspect ratio than the rest of the movie that lasts the entirety of the 5 minute cropdetect window.  Extending how long crop detect goes on for seems like a waste.

I'm debating how to handle this edge case.  Two crop detects at different time stamps?  Pick the one with the larger canvas?  Each crop detect at a dynamic start time depending on the length of the video?