PersistGS: Differentiable Physics for Object Permanence in 4D Gaussian Splatting
Adrian Ramlal, John S. Zelek · 2026
Read PersistGS: Differentiable Physics for Object Permanence in 4D Gaussian Splatting by Adrian Ramlal, John S. Zelek (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-03479.txt.
Description
Ramlal and Zelek keep moving objects alive in 4D scenes while no camera can see them. Dynamic Gaussian Splatting rebuilds scenes from synced multi-camera video. Once an object hides from every training camera, its signal vanishes. Its Gaussians then get no updates and decay. Past fixes lean on learned priors that favor looks over physical truth.
PersistGS couples a differentiable rigid body simulator with Gaussian Splatting instead. The scene splits into per-object Gaussians and collision meshes. Friction and velocity are fit from the path seen before the object hid. The simulator then rolls that path through the hidden stretch. Bounces, friction slowdowns, and turns come out right, unlike straight-line guesses. A centroid silhouette loss cuts trajectory error by 40 percent versus photometric loss. On synthetic scenes, PersistGS beats constant velocity by 2.46 dB PSNR. It lands within 0.19 dB of a ground-truth path bound.
Catalog record
| CREATOR | Adrian Ramlal, John S. Zelek |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 31 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.03479