Support cropping / region of interest

[Breakthrough]

Processing only a sub-set of a video frame should be possible. This can greatly improve performance and detection accuracy in some cases.

To avoid any issues with cropping or output, both save-images and split-video should still output the full video. Cropping of the output can be considered in the future, but should be avoided if possible and done by post-processing the output of PySceneDetect. The same applies to non-rectangular ROIs or overlapping regions.

This was originally brought up in #147. This is also a feature in PyVfxShotDetect [1], so we can consider implementing something similar by creating a wrapper for VideoStream.

[1] https://github.com/ThomasWeckenmann/PyVfxShotDetect/blob/master/scenedetect/bbox_video_stream.py

[ankandrew]

Looking forward this feature 🙏

[cdgraff]

+1 will be amazing have this feature! I need too...

[Breakthrough]

Can anyone share if they require a GUI for this purpose, or is command line acceptable? cc @ankandrew @cdgraff

I'm thinking the regions can be a set of polygon masks similar to how DVR-Scan handles this. Similarily, if we need a GUI, I may repurpose the region editor from DVR-Scan for this purpose, since that has most of the required functionality. This also means we can support non-rectangular masks.

Regardless we should ship something soon and leave the GUI for later.