Title: RU4D-SLAM: Leveraging Uncertainty Reweighting in Gaussian Splatting SLAM for 4D Scene Reconstruction

Abstract: The integration of Simultaneous Localization and Mapping (SLAM) with 3D Gaussian splatting has become increasingly prevalent, facilitating continuous 3D environment reconstruction during movement. Nevertheless, current techniques face significant difficulties in dynamic settings, where the presence of moving objects complicates reconstruction and undermines tracking reliability. While 4D reconstruction—specifically 4D Gaussian splatting—presents a viable solution to these issues, its application within 4D-aware SLAM systems remains largely unexamined. To address this gap, we introduce RU4D-SLAM (Reweighting Uncertainty in Gaussian Splatting SLAM), a robust and efficient framework designed for 4D scene reconstruction. This approach incorporates temporal dimensions into spatial 3D representations and integrates uncertainty-aware perception to manage scene changes, synthesize blurred images, and reconstruct dynamic scenes. We refine dynamic scene representation through the inclusion of motion blur rendering and bolster uncertainty-aware tracking by adapting per-pixel uncertainty modeling, originally intended for static contexts, to cope with blurred imagery. Additionally, we develop a semantic-guided reweighting mechanism for per-pixel uncertainty estimation in dynamic environments and employ a learnable opacity weight to facilitate adaptive 4D mapping. Comprehensive experiments conducted on standard benchmarks reveal that our method significantly surpasses state-of-the-art techniques in both trajectory accuracy and 4D scene reconstruction quality, especially within dynamic scenarios involving moving objects and low-quality input data. Code available: https://ru4d-slam.github.io

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC