// the find
spla-tam/SplaTAM
SplaTAM: Splat, Track & Map 3D Gaussians for Dense RGB-D SLAM (CVPR 2024)
SplaTAM is a dense RGB-D SLAM system that uses 3D Gaussian Splatting as its map representation instead of NeRFs or point clouds. It simultaneously tracks camera pose and builds a photorealistic 3D map from depth+color frames. Aimed at robotics researchers and CV practitioners who need real-time dense mapping with high-quality novel view synthesis as a side effect.
The core idea is genuinely clever — Gaussians are differentiable and fast to render, so pose optimization and map refinement share the same gradient pipeline rather than being separate stages. iPhone live demo via NeRFCapture is a low-friction way to actually test it on real hardware without a dedicated depth sensor. Config files are plain Python dicts, not YAML buried in a framework, which makes reproducing paper results and tweaking hyperparameters straightforward. PLY export means reconstructions plug directly into standard 3DGS viewers like SuperSplat.
Pinned to PyTorch 1.12 / CUDA 11.6 despite claiming broader compatibility — the custom CUDA rasterizer (a git submodule) is what actually dictates this, and recompiling it against newer toolchains is not documented. No incremental map updates: Gaussians are densified globally, so memory and per-frame cost grow with scene size and there's no pruning strategy described for long trajectories. The repo has been dormant since mid-2024 with a 'stay tuned for a faster variant' note that never materialized, so you're adopting research code that is effectively frozen. Wandb is effectively required for benchmarking — there's no fallback logging path described if you don't have an account.