ChronoFuseGS: Multi-Temporal Gaussian Fusion with Per-Splat Persistence and Change Visualization

arXiv cs.GR details ChronoFuseGS, a multi-temporal Gaussian splatting approach for scenes that change between capture sessions. Instead of treating each timestep as a separate reconstruction, it merges multiple independently trained models into a single combined scene, letting Gaussians from one capture contribute to other timesteps when they represent persistent structure.

That matters for anyone building reconstruction pipelines or tools for long-running environments, because the model is designed to extend incrementally as new timesteps arrive without throwing away earlier work. Each Gaussian primitive carries persistence information, so the system can preserve stable geometry while still accounting for seasonal shifts, flooding, snow cover, or other localized changes.

The paper also pairs the reconstruction with a change-aware visualization layer that keeps persistent areas in color and emphasizes what changed across a user-selected time range. Because the persistence is tracked at the Gaussian level, the visualization can expose sub-object changes rather than only coarse object-level differences, which is useful for inspection, editing, and analysis workflows.

The method was evaluated on a real outdoor flood-management dataset captured over seven months across eight recording days, and the authors say the merged model outperforms single-timestep reconstructions in novel-view synthesis while recovering details missing from individual captures. They also released the dataset publicly, which should make the work easier to reproduce and build on.