3D Scene Change Modeling With Consistent Multi-View Aggregation

TL;DR

SCaR-3D tackles object-level 3D scene change detection from dense pre-change and sparse post-change views while enabling continual reconstruction. It fuses a signed-distance-based 2D difference module with a 3D difference aggregation framework that employs multi-view voting, pruning, and segmentation validation within the 3D Gaussian Splatting representation. The work introduces CCS3D for controlled evaluations and demonstrates state-of-the-art precision, IoU, and reconstruction quality across synthetic and real datasets, along with rigorous ablations. The approach supports targeted updates to dynamic regions, reducing drift in unchanged areas and offering practical benefits for scalable scene monitoring, editing, and long-term reconstruction.

Abstract

Change detection plays a vital role in scene monitoring, exploration, and continual reconstruction. Existing 3D change detection methods often exhibit spatial inconsistency in the detected changes and fail to explicitly separate pre- and post-change states. To address these limitations, we propose SCaR-3D, a novel 3D scene change detection framework that identifies object-level changes from a dense-view pre-change image sequence and sparse-view post-change images. Our approach consists of a signed-distance-based 2D differencing module followed by multi-view aggregation with voting and pruning, leveraging the consistent nature of 3DGS to robustly separate pre- and post-change states. We further develop a continual scene reconstruction strategy that selectively updates dynamic regions while preserving the unchanged areas. We also contribute CCS3D, a challenging synthetic dataset that allows flexible combinations of 3D change types to support controlled evaluations. Extensive experiments demonstrate that our method achieves both high accuracy and efficiency, outperforming existing methods.

Figures (6)

• Figure 1: We propose SCaR-3D, a 3D scene change modeling framework that detects changes from dense-view pre-change images and sparse-view post-change images, while seamlessly reconstructing the post-change scene. SCaR-3D significantly outperforms existing 3D change detection methods in change mask accuracy and computational efficiency, and delivers high-quality continual reconstructions.
• Figure 2: Overview of SCaR-3D. We first employ COLMAP for image registration, producing paired pre-change renders and post-change captures. In the 2D Difference Generation stage, features are extracted and a signed distance metric is applied to separate the change regions into two sets. After that, the 3D Difference Aggregation stage integrates multi-view differences through voting, pruning, and segmentation validation. Finally, the change masks are applied to the reconstruction process to update the 3D scene selectively.
• Figure 3: Multi-view pruning. Left: When objects exist (top), masks are scattered and few Gaussians are removed; when objects are removed (bottom), background tracing leads to extensive pruning. Right: Novel-view 3D difference visualization for (b) single-view weighting in \ref{['eq:trace_sing']}, (c) multi-view voting in \ref{['eq:trace_seen']}, and (d) multi-view pruning.
• Figure 4: Qualitative change detection results on the CCS3D dataset. Each pair of rows corresponds to a single scene captured from different viewpoints. The last column, labeled Reference, shows the post-change images from the matched viewpoints.
• Figure 5: Qualitative reconstruction results on novel views. Each pair of rows corresponds to a single scene captured from different novel viewpoints. The changed regions are highlighted with red boxes.