Can triangles beat Gaussians for fast 3D scene mapping?

Standard SLAM uses point clouds or implicit representations; this work reconstructs scenes as differentiable triangle meshes optimized during camera tracking. On Replica and TUM-RGBD benchmarks, it matches competing trackers while producing sharper 3D geometry, and uniquely enables live mesh editing and collision checking—features impossible with Gaussian splatting. The system converts unstructured triangles into connected meshes on-the-fly.