slaqr-mel/mpdecimate_trim

trim video clips based on mpdecimate output, keep audio synced

Script to trim similar/duplicate fragments from video clips. While keeping audio in sync.

Note that it does not do what is often called "dropping frames" (i.e. removing them from container by replacing with PTS of a "similar enough" one from another part of the clip). It actually gets rid of them completely, making the resulting clip shorter in time.

Note also that some variables, such as mpdecimate thresholds or output codec settings, are currently hardcoded in the script. Mostly due to laziness ;-).

Usage

Needs Python3.7+ and ffmpeg.

$ mpdecimate_trim.py [--skip SKIP] [--keep] [--vaapi <render_device_filepath>] [--vaapi-decimate [render_device_filepath]] <filepath>

This will take file at <filepath>, detect frames with certain similarity, re-encode it with them removed (using libx265/hevc_vaapi) and delete the original file.

The --keep switch makes it keep the original.

By default, re-encode happens even if no fragments to trim are found. This can be adjusted by setting --skip to minimum amount of remaining clip parts (e.g. <=1 is equivalent to default, 2 means 1 trimmed fragment, and so on).

The --vaapi option enables VA-API based hardware accelerated transcoding. Note that the script does not check whether supplied input and/or available GPU are capable of performing the transcode, if they are not the process will fail.

The --vaapi-decimate option enables VA-API based hardware accelerated decimate filter. If the optional device path is supplied, this device will be used. Otherwise, it will attempt to use device specified with --vaapi option. If neither device is specified, the script will fail. Note that on some older versions of ffmpeg this might fail even if VA-API transcoding works, not sure why. I have only tested this with >=4.4.1.

vs_decimate?

Was a different experiment, using vapoursynth. Abandoned, because its' decimation algorithm does not fit my needs, and the whole process is also noticeably slower.

In case you want to use it for something, it needs Python3.6+ and vapoursynth with ffms2 and damb.