ChronoGS: Disentangling Invariants and Changes in Multi-Period Scenes

TL;DR

ChronoGS introduces a unified anchor scaffold and temporally modulated Gaussian representation to reconstruct multi-period scenes with non-continuous geometry and appearance changes. By disentangling invariant structure from period-specific variations and employing a geometry-activation mechanism, ChronoGS achieves temporally consistent reconstructions across disparate time spans. The ChronoScene dataset provides a challenging benchmark combining geometric and appearance evolution, and experiments show superior reconstruction quality, temporal fidelity, and efficiency compared with static and continuous-dynamics baselines. The work offers a practical foundation for long-term 3D scene understanding and paves the way for future multi-period, geometry-aware rendering and analysis.

Abstract

Multi-period image collections are common in real-world applications. Cities are re-scanned for mapping, construction sites are revisited for progress tracking, and natural regions are monitored for environmental change. Such data form multi-period scenes, where geometry and appearance evolve. Reconstructing such scenes is an important yet underexplored problem. Existing pipelines rely on incompatible assumptions: static and in-the-wild methods enforce a single geometry, while dynamic ones assume smooth motion, both failing under long-term, discontinuous changes. To solve this problem, we introduce ChronoGS, a temporally modulated Gaussian representation that reconstructs all periods within a unified anchor scaffold. It's also designed to disentangle stable and evolving components, achieving temporally consistent reconstruction of multi-period scenes. To catalyze relevant research, we release ChronoScene dataset, a benchmark of real and synthetic multi-period scenes, capturing geometric and appearance variation. Experiments demonstrate that ChronoGS consistently outperforms baselines in reconstruction quality and temporal consistency. Our code and the ChronoScene dataset are publicly available at https://github.com/ZhongtaoWang/ChronoGS.

Figures (22)

• Figure 1: As time goes by, scenes naturally evolve — sunlight shifts, new structures emerge, seasons change, and so on. Cross-period captures therefore exhibit variations in both geometry and appearance. The red and blue overlays in the illustration simply exemplify the differences between the two periods. These coupled changes pose significant challenges for existing methods. Our approach, ChronoGS, robustly reconstructs and disentangles both geometry and appearance variations, delivering consistent and faithful results across time.
• Figure 2: For a given camera $C_j^t$, we select visible anchors from our learned anchor scaffold of union geometry across periods. We use per-anchor's features and global features to render the multi-period scene. Temporal behavior is modeled via the anchor's local period-varying feature $f_i^{\text{var}}(t)$ and global period-varying feature $g(t)$ modulated by temporal encoding $e(t)$. After concatenating them with anchor's period-invariant base feature $f_i^{\text{base}}$, a lightweight MLP predicts Gaussian attributes $\{\alpha, S ,c\}$ . These attributes, along with the learned Gaussian position offset $\mu$, produce a small cluster of Gaussians per anchor. We render the Gaussians generated by all visible anchors via differentiable splatting with alpha blending to obtain the image $\hat{I}_j^t$, and optimize with photometric losses between the ground truth image. This design enables consistent reconstruction of both geometry and appearance evolution on multi-period scenes.
• Figure 3: Illustration of the temporal geometry activation mechanism.(a). Anchor scaffold encodes the union of geometry across all periods, and generated gaussians with negative opacity are deactivated at certain periods, enabling the model to adaptively represent geometry variations over time. (b). Comparison between Scaffold-GS and our ChronoGS trained on all periods' images. Scaffold-GS entangles structures from different periods, while ChronoGS successfully distinguishes and reconstructs period-specific geometry at $\text{Period}_0$ and $\text{Period}_1$.
• Figure 4: Comparison between Scaffold-GS trained in one period (left) and ChronoGS trained across all periods (right). Regions highlighted in red show that ChronoGS benefits from multi-period joint training, leveraging more observations to recover more accurate geometry and produce more complete reconstructions.
• Figure 5: (a). Examples from the ChronoScene dataset. Top: real-world aerial captures showing natural evolution across long temporal intervals (e.g., construction and season change). Bottom: our synthetic scenes edited to simulate realistic changes with controllable geometry and appearance changes. (b). Number distribution of image and point cloud across periods for 6 real-world and 6 synthetic scenes. Left bars show image counts, right bars show sparse point cloud counts.