Learning from Automatic 3D Gaussian Splatting Edits for Online Bayesian Scene Change Detection

Learning from Automatic 3D Gaussian Splatting Edits for Online Bayesian Scene Change Detection. By Changrui Zhu and one other, from University College London. Published on the twenty-first of October. Its code repository is worked out: about four hundred source files, with a readme and a licence.

Online scene change detection means taking observations as they arrive and deciding which regions of a scene have changed relative to a reference. The authors frame this as updating beliefs about change with each new image, through generalised Bayesian evidence fusion. That framing has three obstacles. The prior probability of change is not known beforehand. Neither are the evidence models. And numerical Bayesian inference can be costly.

Their answer is to learn what is missing offline, from automatic edits made to a reference 3D Gaussian Splatting model. Editing the reference produces labelled training views without any manual annotation. Paired with unchanged reference observations, those edited views supply the supervision needed to learn scene-specific evidence models, and to estimate the effective pixel prior and the response statistics.

At run time, online prediction combines the current evidence with historical support. The belief itself lives in an accumulated 3D Gaussian field, updated by closed-form local Bayes factors. Those factors are parameterized by the response statistics learned offline, so the cost of numerical inference is replaced by a closed-form update. A final step, object-guided refinement, turns the accumulated belief into the change masks.

The method reports state-of-the-art accuracy on two datasets. On PASLCD it reaches sixty-seven point six four percent mIoU. On CCS3D it reaches seventy-five point eight two percent.

Speed is reported on PASLCD as well. The complete pipeline runs at seventeen point two frames per second over complete query sequences, the fastest among the online methods evaluated.

All three unknowns, the prior, the evidence model and the expensive inference, are handled by the same move. Edits to the reference scene generate their own labels, and what is learned from them makes the online Bayesian update both cheap and closed-form.