Topic: Segmentation
What about applying a real-time segmentation algorithm on the motion vectors to find the different "objects of the scene" so we could blend their motion vectors one object at time without impacting near objects (number of objects handled would be a setting).
Maybe using different motion vector sizes at same time would help to catch local and more global movements.
A Other idea would be to use un-linear motion vectors like cubic, spline...
I would like to know what you guys think theses, what would be the cost and gain to implement thems etc.
Great work !