4 ms·
> trained only on Text-Image pairs and unlabeled videos This is fascinating. It's able to pick up sufficiently on the fundamentals of 3D motion from 2D videos
by radarsat1 4y ago
> trained only on Text-Image pairs and unlabeled videos
This is fascinating. It's able to pick up sufficiently on the fundamentals of 3D motion from 2D videos, while only needing static images with descriptions to infer semantics.