Online scene change detection asks a system to decide which parts of a scene have changed relative to a known reference, while new observations keep arriving. The challenge is to integrate each new image without revisiting the whole history. The authors frame this as a Bayesian inference problem, maintaining a belief about changed regions and updating it as evidence accumulates.
The proposed method uses generalized Bayesian evidence to perform these updates. The title indicates that the reference is represented as a 3D Gaussian splatting, and the system learns from automatic edits to that representation. This connects a modern neural rendering format to a classical probabilistic framework for change detection.
Because the update is online, the method is suited to settings where images arrive sequentially and decisions must be made promptly. The abstract focuses on the framework rather than reporting specific results, but the contribution is the combination of Bayesian evidence with 3D Gaussian splatting edits for continuous scene monitoring.