NICE-SLAM: Neural Implicit Scalable Encoding for SLAM

Zhu et al. present NICE-SLAM, a dense online SLAM system that reconstructs an implicit neural scene representation while simultaneously tracking the camera. It targets a gap left by iMAP, the first neural-implicit SLAM system: a single shared MLP over the entire scene is convenient for small rooms but cannot scale to larger spaces because every observation perturbs every parameter, producing over-smoothing and forgetting.

NICE-SLAM replaces the monolithic MLP with a hierarchical feature grid. Three voxel grids at coarse, mid, and fine resolutions store local geometric features, and a small shared decoder MLP turns trilinearly interpolated features into occupancy and color. Because each observation only updates the small subset of grid cells visible from the current camera pose, the system avoids the global forgetting of iMAP and scales to multi-room indoor environments while keeping runtime roughly constant in scene size. Pre-trained geometric priors on the decoder MLPs let the model fill in unobserved regions plausibly, yielding completed reconstructions even from sparse views.

The mapping and tracking threads are decoupled and run alternately: mapping fixes the camera pose and updates the scene grids on a sliding window of keyframes; tracking fixes the scene and refines the latest camera pose by minimizing the same RGB-D photometric and geometric losses. On five challenging datasets including Replica, ScanNet, and TUM-RGBD the system achieves competitive mapping and tracking accuracy versus iMAP and classical TSDF-fusion baselines, while producing significantly more detailed reconstructions on large scenes.

NICE-SLAM became the reference architecture for the wave of neural- implicit SLAM follow-ups (Vox-Fusion, Co-SLAM, Point-SLAM) and is the counterpart to DROID-SLAM (Teed and Deng, 2021) within this corpus — both attack online SLAM with neural representations but from opposite ends of the explicit-implicit spectrum.