Abstract
Dense simultaneous localization and mapping (SLAM) is pivotal for embodiedscene understanding. Recent work has shown that 3D Gaussians enablehigh-quality reconstruction and real-time rendering of scenes using multipleposed cameras. In this light, we show for the first time that representing ascene by 3D Gaussians can enable dense SLAM using a single unposed monocularRGB-D camera. Our method, SplaTAM, addresses the limitations of prior radiancefield-based representations, including fast rendering and optimization, theability to determine if areas have been previously mapped, and structured mapexpansion by adding more Gaussians. We employ an online tracking and mappingpipeline while tailoring it to specifically use an underlying Gaussianrepresentation and silhouette-guided optimization via differentiable rendering.Extensive experiments show that SplaTAM achieves up to 2X state-of-the-artperformance in camera pose estimation, map construction, and novel-viewsynthesis, demonstrating its superiority over existing approaches, whileallowing real-time rendering of a high-resolution dense 3D map.