Inconsistent video splitting about pyscenedetect

[mohannad57]

I'm not sure about the inconsistency but I know that on version 0.4 this was changed but the documentations are still.

"•specifying -o OUTPUT_FILE.mkv will now automatically split the input video, generating a new video clip for each detected scene in sequence, starting with OUTPUT_FILE-001.mkv
•CSV file output is now specified with the -co / --csv-output option (note, used to be -o in versions of PySceneDetect < 0.4)"
source: https://github.com/Breakthrough/PySceneDetect/releases

[alegaballo]

Oh okay, I didn't notice that, thank you. Now the .csv file is correct but still the number of scenes if I try to use -o scene.mkv is different

[mohannad57]

That part works for me just fine. so I will leave it for the experts as I'm very new to all this. :) sorry

[Breakthrough]

Hmm interesting... Using the new command-line options (-o for the video output, -co for the CSV output), could you provide the full output you get now? Of the scenes that are written to disk, are they cut at the appropriate place (near where the timecode in the CSV output is)?

[alegaballo]

Sorry for being this late, I was quite busy. Try with this video with the default configuration:
scenedetect -i video.mp4 -c scenes.csv -o out.mp4
the number of scenes is correct but their duration is different from the one described in the csv

[Breakthrough]

Hello all;

This has been fixed in the most recent version of PySceneDetect (v0.5-beta). The stable v0.5 release should be out sometime this month.

The issue is that when just splitting videos without re-encoding, they can only be split in certain cases on keyframe boundaries. This also means that sometimes, when several scenes are very small and close together, they may be concatenated together, or not generated properly.

To fix this, I have integrated PySceneDetect v0.5 with ffmpeg to allow re-encoding the output video. To accomplish this in v0.5, you would use the split-video command with the -p/--precise option:

scenedetect -i input_video.mp4 -o output_dir detect-content split-video -p

By default, split-video will use ffmpeg in stream copy mode, similar to how mkvmerge works (e.g. how PySceneDetect v0.4 output scenes). If ffmpeg is not available, PySceneDetect will then attempt to use mkvmerge if possible; you can force the use of mkvmerge using the -m/--mkvmerge option with the split-video command.

Note that you can also save thumbnail images now via the save-images command.

scenedetect -i input_video.mp4 -o output_dir detect-content split-video -p save-images

I have just tested this on v0.5, and indeed, some output files were unreadable without either the -p (precise/re-encode) or -m (use mkvmerge) flags. Obviously using the -p flag is much slower than just copying the streams, but it is significantly more accurate in terms of where the output cuts are placed. You can also override the codec options, e.g.:

scenedetect -i input_video.mp4 -o output_dir detect-content split-video -f "-c:v libx264 -c:a aac"

Feel free to test the v0.5-beta if you have a chance and provide some feedback. If you find any issues in the new version using the -p flag, by all means, please bring it to my attention by opening a new issue.

Thank you for reporting this, and the feedback you both (@alegaballo @mohannad57) provided; feel free to share your thoughts on the new command line interface in the v0.5-beta as well.