That's a great question and a very good intuition.

We explored that idea initially, but decided to move away from relying on scenes for deduplication for 2 reasons:

1. algorithmic scene segmentation is in general less accurate than shot segmentation.

2. shots within the same scene can be drastically different visually. I.e. it's possible to throw out legitimate shots that otherwise would've been perfect for matching. An obvious example of this is when a scene already contains match cuts.