Gaussian Splatting SLAM
Uses Gaussians as the sole map representation for monocular SLAM, jointly optimizing poses and primitives for incremental reconstruction and NVS.
Authors / Team
Hidenobu Matsuki · Researcher
Year
2024
Deep Dive
The work brings 3DGS into online SLAM: instead of assuming offline SfM poses, it formulates direct photometric camera tracking against the evolving Gaussian map and adds geometric verification leveraging explicit primitives for incremental dense mapping. It targets monocular live operation at interactive rates and extends to RGB-D when depth is available. The MonoGS codebase, project page, and video provide reproducibility and qualitative demos on challenging indoor sequences.
What we learn
- 01
Unifying tracking and mapping in one explicit representation reduces cross-module representation mismatch.
- 02
Incremental dense mapping still needs geometric checks; photometric consistency alone does not remove monocular ambiguities.
Verbatim quote
"We present the first application of 3D Gaussian Splatting in monocular SLAM, the most fundamental but the hardest setup for Visual SLAM."— source ↗
Tags
Links
Sources