// the find
muskie82/MonoGS
[CVPR'24 Highlight & Best Demo Award] Gaussian Splatting SLAM
MonoGS is a dense SLAM system that represents the environment as 3D Gaussian splats instead of the usual point clouds or NeRFs. It runs from a single monocular camera (no depth sensor required), which is the hard version of this problem. Aimed at robotics researchers and computer vision practitioners who want a photorealistic map alongside pose estimates.
Monocular-only mode is the real contribution — most dense SLAM systems cheat with depth sensors; this one optimizes camera pose jointly with the Gaussian map. The frontend/backend split into separate processes (slam_frontend.py / slam_backend.py) is a sensible architecture for real-time systems where tracking latency matters. Live Realsense support with a config file means you can actually run it on hardware, not just datasets. The speedup branch reportedly hits 10fps on fr3/office, which is competitive for a method doing full 3DGS rendering in the loop.
Pinned to PyTorch 1.12 and CUDA 11.x — that's two major PyTorch versions behind and will fight with any modern environment; don't expect a clean conda install without version archaeology. Last commit was August 2024 and the speedup branch was never merged to main, so you're getting a research snapshot, not a maintained library. Multi-process reproducibility is explicitly disclaimed in the README, which is a red flag for anyone trying to benchmark or compare results. Stereo support is labeled 'experimental' and there's only one EuRoC sequence config, so don't build on that path.