Object-focused motion representations introduced for video
The author argues that current video representations mix motion across an entire scene, and introduces a “localized context-aware” approach aimed at individual objects.
TLDR
Real-world scenes have many objects moving at once. A post argues that current video representations entangle that motion globally—an obstacle to understanding and generating motion for individual objects. It introduces “localized context-aware motion representations” to address that problem.
Combined views
46
1 Source, first seen 17d ago