Title: MipSLAM: Anti-Aliased Gaussian Splatting for Simultaneous Localization and Mapping

Abstract:

This study presents MipSLAM, a novel SLAM framework built on frequency-aware 3D Gaussian Splatting (3DGS). The system is designed to deliver robust pose estimation and high-fidelity, anti-aliased novel view synthesis, even when camera configurations vary. Current SLAM approaches based on 3DGS frequently encounter issues such as trajectory drift and aliasing artifacts, largely stemming from insufficient filtering mechanisms and reliance on purely spatial optimization techniques.

To address these challenges, we introduce an Elliptical Adaptive Anti-aliasing (EAA) algorithm. This method estimates Gaussian contributions through geometry-aware numerical integration, thereby bypassing the need for expensive analytic calculations. Additionally, we propose a Spectral-Aware Pose Graph Optimization (SA-PGO) module. This component reimagines trajectory estimation within the frequency domain, utilizing graph Laplacian analysis to effectively mitigate high-frequency noise and drift.

Comprehensive testing on the TUM and Replica datasets reveals that MipSLAM sets a new state-of-the-art standard for both localization precision and rendering quality across various resolutions. The source code is publicly accessible at https://github.com/yzli1998/MipSLAM.

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