Computer vision · Public code / Dynamic scene reconstruction
Decoding 4D Gaussians
How can a single moving camera describe a scene that is also changing?
The work
The repository connects a frozen SpaTrackerV2 model to trainable decoder heads. They turn features and 3D point maps from a monocular video into a set of Gaussians that can change over time.
View repository on GitHubTry the idea
Look at the same synthetic ribbon from two angles. Switch to point centers, then restore the colored footprints. Scrub time while holding the camera still to see which movement comes from the scene.
Sculpt a moving scene