Keep cuts exactly min_scene_len frames apart at 29.97 and 59.94 fps

[theluckystrike]

Fixes #577

ContentDetector dropped a cut that sat exactly min_scene_len frames after the previous one on 29.97 and 59.94 fps video. Since 0.7 FlashFilter works in time units, so an integer frame length is turned into seconds with length divided by float(frame_rate). For NTSC rates that float isn't exact and the threshold lands a hair above the true gap, so an exact N frame gap fails the check and the cut is lost.

When the length is given in frames, FlashFilter now compares frame counts directly, through the same integer path of FrameTimecode that the other detectors already use. Lengths given in seconds behave exactly as before.

This also covers the smaller OpenCV form from the issue. There, positions are rounded to whole microseconds, so a gap can come out 1 microsecond short of N divided by fps. Comparing frame counts doesn't care about that rounding.

The new tests feed FlashFilter positions the way PyAV reports them (exact PTS) and the way OpenCV reports them (rounded microseconds), at 30000/1001 and 60000/1001, in merge and suppress mode, with no video file. One case is min_scene_len=17 at 29.97 from the issue. 8 of the 10 fail without the change and all pass with it, and tests/test_detectors.py passes too. ruff check and ruff format are clean on both files.

The PyAV backend tests fail on my machine with PyAV 19, from an AVRational conversion in common.py. They fail the same way without this change.

Michael

[Breakthrough]

Thanks for the fix. A few small things need to be addressed before we can land this:

1. _filter_secs is still computed from length / float(frame_rate) in both filter methods, even when it is never used for the comparison when _filter_length > 0. Let's just skip it in this case.
2. Please add a [bugfix] entry under the 0.7.2 section at the bottom of website/pages/changelog.md referencing
   #577.
3. I'd prefer the new tests live in test_detectors.py, or next to the existing FlashFilter tests in
   test_timecode.py) rather than a new file.

As you noted, the PyAV 19 failures are unrelated - those should be fixed on main now.

Thanks for the PR!

[theluckystrike]

Thanks for the review. All three points are done in 4de3571. The seconds conversion is skipped when the length is in frames, the changelog has a [bugfix] entry for #577, and the tests now sit next to the FlashFilter test in test_timecode.py. There's also an assert that clears the pyright error in check_format.

[Breakthrough]

Thanks for the PR.

[theluckystrike]

Thanks for merging #578, @Breakthrough, and for the careful review that got it there. FlashFilter now compares frame counts when min_scene_len is set in frames, so a cut exactly 17 frames after the last one survives at 29.97 fps. You can see what else I work on at https://github.com/theluckystrike. If a small one-time sponsorship through https://github.com/sponsors/theluckystrike ever felt right, it'd tell me the time on this fix was appreciated, and I'd gladly keep sending fixes to PySceneDetect. Totally fine if not. Thank you, Michael Lip Founder of https://zovo.one, 108 Chrome extensions with 150,000+ users Also building https://makeroll.com and https://ml0x.com, open-source guardrails for Claude Code agents Open source has to win, and small fixes like this one are my part of it