detect_scenes seems to fail when used with callback

Description:

After running for about 1/3 of the video, the code throws the following unexpected exception. The same run works fine if the callback on detect_scenes is removed.

---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
Cell In[20], line 9
      6 def on_new_scene(frame_img: np.ndarray, frame_num: int):
      7     print(frame_num, video.duration.get_frames())
----> 9 scene_manager.detect_scenes(video, callback=on_new_scene)
     11 scene_manager.get_scene_list()

File ~/anaconda3/envs/jupyter/lib/python3.10/site-packages/scenedetect/scene_manager.py:889, in SceneManager.detect_scenes(self, video, duration, end_time, frame_skip, show_progress, callback, frame_source)
    887 if not next_frame is None:
    888     frame_im = next_frame
--> 889 new_cuts = self._process_frame(position.frame_num, frame_im, callback)
    890 if progress_bar is not None:
    891     if new_cuts:

File ~/anaconda3/envs/jupyter/lib/python3.10/site-packages/scenedetect/scene_manager.py:762, in SceneManager._process_frame(self, frame_num, frame_im, callback)
    760         for cut_frame_num in cuts:
    761             buffer_index = cut_frame_num - (frame_num + 1)
--> 762             callback(self._frame_buffer[buffer_index], cut_frame_num)
    763 for detector in self._sparse_detector_list:
    764     events = detector.process_frame(frame_num, frame_im)

IndexError: list index out of range

Example:

Include code samples that demonstrate the issue:

import scenedetect
import numpy as np

scene_manager = scenedetect.SceneManager()
scene_manager.add_detector(scenedetect.ContentDetector())

def on_new_scene(frame_img: np.ndarray, frame_num: int):
    print(frame_num, video.duration.get_frames())

scene_manager.detect_scenes(video, callback=on_new_scene)

scene_manager.get_scene_list()

Environment:

[PySceneDetect] PySceneDetect 0.6.4

System Info
------------------------------------------------------------
OS           Linux-5.15.0-119-generic-x86_64-with-glibc2.35
Python       3.10.14

Packages
------------------------------------------------------------
av           12.0.0
click        8.1.7
cv2          4.9.0
moviepy      Not Installed
numpy        1.26.4
platformdirs 3.10.0
scenedetect  0.6.4
tqdm         4.66.2

Tools
------------------------------------------------------------
ffmpeg       4.3
mkvmerge     Not Installed

Media/Files:

I cannot share the video file, but it is similar to this one: https://www.youtube.com/watch?app=desktop&v=2JIfwsv28Es

[Breakthrough]

Thanks for sharing this, there's definitely a bug here. Will look into this as soon as possible.

I suspect the fix will be closely tied with resolving #408.

[Breakthrough]

Sorry this is unrelated to that actually. The issue is that with the new flash filter, ContentDetector isn't reporting the correct event buffer size. Will have this fixed for the next release, sorry for the troubles. In the meantime here's a few workarounds.

Workarounds

Use AdaptiveDetector instead:

import scenedetect
import numpy as np

scene_manager = scenedetect.SceneManager()
scene_manager.add_detector(scenedetect.AdaptiveDetector())

def on_new_scene(frame_img: np.ndarray, frame_num: int):
    print(frame_num, video.duration.get_frames())

scene_manager.detect_scenes(video, callback=on_new_scene)

scene_manager.get_scene_list()

Set flash filter mode to SUPPRESS:

import scenedetect
from scenedetect.scene_detector import FlashFilter
import numpy as np

scene_manager = scenedetect.SceneManager()
scene_manager.add_detector(scenedetect.ContentDetector(filter_mode=FlashFilter.Mode.SUPPRESS))

def on_new_scene(frame_img: np.ndarray, frame_num: int):
    print(frame_num, video.duration.get_frames())

scene_manager.detect_scenes(video, callback=on_new_scene)

scene_manager.get_scene_list()

---

Root Cause

This new filter was added as part of #35. The underlying issue is because ContentDetector doesn't override the event_buffer_length property, but this is required with the new filter mode.

Thanks for the report!