Gaussian-SLAM

Why we included this project

Gaussian-SLAM consumes a live RGB-D stream and represents the scanned environment as a set of 3D Gaussians rather than the usual point cloud or mesh. That pays off when the map has to double as imagery: the dense reconstruction can be rendered from new angles with realistic detail, which suits robotic perception, AR navigation, and photorealistic scene capture. The authors validate it on Replica, TUM RGB-D, and ScanNet, the standard SLAM benchmarks, so results are easy to line up against other dense pipelines. This is research code, not a product: expect to set up the conda environment yourself, run it on a CUDA GPU (RTX 3090 or A6000 in their tests), and edit config files. Teams deciding whether gaussian-splatting-based tracking is worth adopting will find it a useful baseline.