SPLAT: Streaming Perception of Language-Aligned Temporal Gaussians
A unified framework for real-time 4D scene perception with dynamic reconstruction, explicit motion modeling, and open-vocabulary understanding from unposed video streams.
We present SPLAT, a feed-forward framework for streaming 4D scene perception that unifies dynamic geometry reconstruction, explicit motion modeling, and open-vocabulary semantic understanding within a single representation of language-aligned temporal Gaussians.