Point-SLAM: Dense Neural Point Cloud-based SLAM

Sandström and colleagues present a dense neural SLAM system that anchors the features of a neural scene representation in a PointCloud rather than the sparse voxel or hash grids favoured by contemporaneous neural SLAM work. The point set is grown incrementally as new RGBD frames arrive, with anchor density adapted to the local information density of the input — flat walls receive few points, while textured or geometrically complex regions are densified.

Both the mapping and tracking stages share the same point-based field. The authors minimise an RGBD-based re-rendering loss that compares synthesised colour and depth against the incoming sensor stream; the same loss drives gradient updates to the per-point feature vectors and to the current camera pose. Because anchors are sparse where they need to be sparse, the system spends less compute on empty space than grid-bound competitors and is able to allocate finer detail near surfaces.

Experiments on Replica, TUM-RGBD, and ScanNet report tracking accuracy and rendering quality (PSNR, SSIM, LPIPS) on par with or better than prior dense neural SLAM methods. The data-driven anchor placement also reduces memory relative to fixed-resolution grid encodings such as those used by iMAP and NICE-SLAM, making the approach a useful reference point for later GaussianSplatting-based SLAM efforts (Gaussian Splatting SLAM, GS-SLAM) that retain the explicit-primitive philosophy but swap point anchors for anisotropic Gaussian primitives. A reliability check on new candidate points prevents the cloud from absorbing every transient observation, keeping the active point set bounded even under repeated revisits, which is one of the more practical contributions of the work to incremental neural mapping. Source code is released and the design remains a common citation for input-adaptive neural map representations in real-time SLAM.