Ability to keep the most significant video frames in dataset

[zhiltsov-max]

A video can contain lots of frames, most of which do not bring any new information. A dataset created from a video can be significantly reduced without affecting model performance, if we keep only the most significant frames.

References:

• BubbleNets: https://arxiv.org/pdf/1903.11779.pdf, https://github.com/griffbr/BubbleNets
• Frames can be checked for similarity (similarly to near duplicate removal)