Modern 4D reconstruction models can track millions of 3D points, but they still lack object-level understanding.
From static scenes to dynamic worlds—Models see points, but not objects.
- Modern 4D reconstruction models can track millions of 3D points over time, but they lack object-level understanding.
- These methods represent scenes as dense point trajectories, without capturing semantic structure or object identity.
- As a result, they fail to model:
- Object permanence across time
- Coherent motion of entities
- Compositional scene structure
Our Goal:
To learn per object 4D representations for dynamic scenes.
Enabling tracking, reasoning, and structured understanding of dynamic environments