UCL paper: auto-edited 3D Gaussian Splatting enables online Bayesian change detection

Changrui Zhu · hf · 2026-10-27

A UCL paper by Changrui Zhu and Simon Julier introduces a new online scene change detection method:

Challenge: change priors and evidence models are unknown upfront, and numerical Bayesian inference is costly;
Approach: apply automatic edits to a reference 3D Gaussian Splatting (3DGS) model, generating labelled training views without manual annotation to learn scene-specific evidence models, effective pixel priors and response statistics offline;
Online: generalized Bayesian evidence fusion updates change beliefs per incoming image, closed-form local Bayes factors update an accumulated 3D Gaussian field, and object-guided refinement produces final change masks;
Achieves state-of-the-art results. Project page: changruizhu96.github.io/work/lego-scd.

More from Research