PySceneDetect v0.6-dev2

[Breakthrough]

Release Notes

This is a pre-release of the upcoming PySceneDetect v0.6. You can grab it from pip:

pip install --pre scenedetect[opencv]

PySceneDetect v0.6 is a major breaking change including better performance, configuration file support, and a more ergonomic API. The new minimum Python version is now 3.6. See the Migration Guide for information on how to port existing applications to the new API. Most users will see performance improvements after updating, and changes to the command-line are not expected to break most workflows.

The main goals of v0.6 are reliability and performance. To achieve this required several breaking changes. The video input API was refactored, and many technical debt items were addressed. This should help the eventual transition to the first planned stable release (v1.0) where the goal is an improved scene detection API.

Changelog

Overview:

• Support for Python 2.7 has been dropped, minimum supported Python version is 3.6
• Support for OpenCV 2.x has been dropped, minimum OpenCV version is 3.x
• Decoding now runs in parallel with detection resulting in significant performance improvements
• Adds support for configuration file via command line or in user settings folder
• Adds support for multiple video backends, PyAV is now supported in addition to OpenCV
• Breaking API changes to VideoManager (replaced with VideoStream), StatsManager, and save_images()
  • See the Migration Guide for details on how to update from v0.5.x

Command-Line Changes:

• Configuration files are now supported via -c/--config or from a user config folder (see documentation for details)
• -i/--input may no longer be specified multiple times (use an external tool like ffmpeg to perform concatenation first)
• -l/--logfile now respects -o/--output
• New -b/--backend option can be set to use a specific video backend
  • Current options are opencv (default) and pyav if the av package is installed
  • Run scenedetect help to see a list of backends available on the current system
• -v/--verbosity now takes precedence over -q/--quiet
• detect-adaptive command:
  • The --drop-short-scenes global option now works correctly with detect-adaptive
• detect-content command:
  • Default threshold -t/--threshold lowered to 27 to be more sensitive to shot changes (#246)
  • Add override for global -m/--min-scene-len option
• detect-threshold command:
  • -p/--min-percent and -b/--block-size have been removed
  • Add override for global -m/--min-scene-len option
• split-video command:
  • Long name for -a has been changed to --args (from --override-args)
  • -c/--copy now uses ffmpeg instead of mkvmerge
  • New -m/--mkvmerge flag can be set to use mkvmerge instead of ffmpeg

API Changes:

• New VideoStream replaces VideoManager and supports both OpenCV (VideoStreamCv2) and PyAV (VideoStreamAv) backends (#213)
  • Improves video seeking invariants, especially around defining what frames 0 and 1 mean for different time properties (frame_number is 1-based whereas position is 0-based to align with PTS)
  • See test_time_invariants in tests/test_video_stream.py as a reference for specific behaviours of these properties, and a test video detailing visually what is expected
  • Both command-line and public-facing API outputs still retain 0-based frame numbers (this will be changed in v1.0)
• Changes to SceneManager:
  • SceneManager is now responsible for frame downscaling (see the downscale and auto_downscale properties)
  • detect_scenes() now performs video decoding in a background thread which greatly improves performance in many cases
  • clear() now also clears any
  • detect_scenes() no longer displays a progress bar by default (set show_progress=True to restore the previous behaviour)detectors, as detectors are stateful
  • get_scene_list() now returns an empty list if there are no detected cuts (previously one scene with the duration of the video was returned)
    • To restore the previous behaviour, specify start_in_scene=True
    • Command-line output is unaffected, and still reports 1 scene spanning the entire video if no cuts were found
• New high-level detect function in scenedetect module that only has two required arguments: the path to a video (path) and a detector (detector)
• save_images() no longer accepts downscale_factor, since there is already the ability to resize images via the scale or height/width arguments
• StatsManager load/save methods now accept a path or an open file handle
• Video splitting functions no longer support multiple input videos for concatenation (scenedetect.video_splitter)
• Fixed issue with previous_frame method of FrameTimecode allowing a negative frame number
• Merged constants MAX_FPS_DELTA and MINIMUM_FRAMES_PER_SECOND_DELTA_FLOAT in scenedetect.frame_timecode into new MAX_FPS_DELTA constant
• video_manager parameter has been removed from the AdaptiveDetector constructor
• split_video_ffmpeg and split_video_mkvmerge functions in scenedetect.video_splitter arguments have been renamed and defaults updated:
  • suppress_output is now show_output, default is False
  • hide_progress is now show_progress, default is False
• block_size argument has been removed from the ThresholdDetector constructor
• calculate_frame_score method of ContentDetector has been renamed to _calculate_frame_score, use new module-level function of the same name instead

Known Issues

• Image sequences or URL inputs are not supported by the save-images or split-video commands
• Variable framerate videos (VFR) are not fully supported, and will yield incorrect timestamps (#168)
• The -l/--add-last-scene option in detect-threshold cannot be disabled
• All other items still open in the [v0.6 Milestone Tracker(https://github.com/Breakthrough/PySceneDetect/milestone/9)

---

This discussion was created from the release PySceneDetect v0.6-dev2.