We propose Slot4D, a framework that combines a frozen VGGT-style backbone with a Slot Attention module to learn object-centric representations in dynamic scenes.
- The VGGT backbone extracts geometry-aware features, including dense 3D point maps and visual embeddings from multi-view inputs.
- A Slot Attention module operates on these features, where a fixed set of slots iteratively attends to and groups pixels/points into object-centric representations.
- Slot assignment is guided by visual similarity and spatial proximity, enabling grouping in 4D (3D space + time).
- Each slot is decoded to reconstruct its contribution to the scene, producing object masks and slot-wise reconstructions.
This design enables consistent object grouping across frames and moves beyond point-based tracking toward structured, object-centric 4D understanding.