Setting different values ​​for min_scene_len results in the same content being divided into different scenes

[iaaidev1]

Using the default value of min_scene_len, the following results are obtained:

[
    ...
    {
        "url": "",
        "scene_detect": {
            "start_frame": 166,
            "start_time": "00:00:08.300",
            "end_frame": 245,
            "end_time": "00:00:12.250"
        }
    },
    {
        "url": "",
        "scene_detect": {
            "start_frame": 246,
            "start_time": "00:00:12.300",
            "end_frame": 295,
            "end_time": "00:00:14.750"
        }
    },
    ...
]
[240, 245] interval split in the previous interval.

min_scene_len is set to 2s (2 * video.frame_rate), and the following results are obtained:

[
    ...
    {
        "url": "",
        "scene_detect": {
            "start_frame": 185,
            "start_time": "00:00:09.250",
            "end_frame": 239,
            "end_time": "00:00:11.950"
        }
    },
    {
        "url": "",
        "scene_detect": {
            "start_frame": 240,
            "start_time": "00:00:12.000",
            "end_frame": 661,
            "end_time": "00:00:33.050"
        }
    },
    ...
]
[240, 245] interval split in the next interval。

video.frame_rate is 20.0

[Breakthrough]

What detector are you using? There is a known issue for AdaptiveDetector in #408.

[iaaidev1]

What detector are you using? There is a known issue for AdaptiveDetector in #408.
this is the code:

content_detector = ContentDetector(threshold=threshold, min_scene_len=min_scene_len)
source_scene_list = detect(
    video_path=video_path,
    detector=content_detector,
    start_time=start_time,
    end_time=end_time,
    show_progress=True)

The problem can be reproduced on versions 0.6.4 and 0.6.5

[Breakthrough]

Any time the min_scene_len parameter is enforced, the results will differ since PySceneDetect needs to determine what to do. In previous versions (v0.6.3 and older), any cut points that occurred before min_scene_len were ignored.

Newer versions (v0.6.4+) allow you to customize what happens with the filter_mode param. By default, ContentDetector now merges cuts shorter than min_scene_len with the preceding scene. This greatly improves performance on videos with lots of flashing.

You can still get the old behaviour by setting the filter_mode param of ContentDetector to FlashFilter.Mode.SUPPRESS, e.g.:

from scenedetect.scene_detector import FlashFilter
content_detector = ContentDetector(..., filter_mode=FlashFilter.Mode.SUPPRESS)

To explain what both modes do:

• FlashFilter.Mode.MERGE: New default for 0.6.4+, merges scenes shorter than min_scene_len with previous scene
• FlashFilter.Mode.SUPPRESS: Previous default before 0.6.4, ignores new scenes until min_scene_len has passed

Let me know if this explains the differences you're seeing!

[iaaidev1]

I understand the problem, thank you very much.