Create mini-scenes for action recognition

[sfmig]

Given:

• a video
• a set of tracked detections (bboxes) per frame
• a list of starting frames and corresponding window widths

Create a set of mini-scenes, that are clips of the cropped individuals at the specified time windows.

It could be useful if the user can select whether the window should be centred or starting at the given frame.

The idea would be to create clips that are useful for training a pixel-based action recognition model (see baboon land project or deepEthogram datasets)