EECS · Neural Implicit SLAM · 2024

Uncertainty-Aware
NeRF-SLAM

Tzu-Chieh TaiElaina MannErika Chen

We incorporate entropy-based uncertainty with covisibility for keyframe selection in neural implicit SLAM, yielding improved localization accuracy over SNI-SLAM while preserving near real-time performance through a fully vectorized batch implementation.

0.54ATE Mean (cm)
0.68ATE RMSE (cm)
1.21Depth L1 (cm)
0.68FPS
↓   scroll to explore
01 · Method

Pipeline overview

H
Uncertainty calc.
Shannon entropy
per keyframe ray
⊞
KF selection
Entropy + covisibility
joint scoring
◈
SNI-SLAM
Mapping & tracking
triplane NeRF
repeats each frame

Entropy visualizer

Rays are cast from a camera into the scene. Each sample point along a ray has a weight — larger rings = higher weight. A concentrated peak = confident surface; a spread distribution = high uncertainty. Drag to rotate.

3D visualization: rays from camera origin, sample points shown as rings scaled by volume rendering weight wi. Red dots mark the true surface position. Adjust the slider to change how peaked or flat the weight distribution is.
— bits · entropy
loading…
Weight distribution spread (uncertain)
Presets
drag to orbit · scroll to zoom
H = −Σᵢ wᵢ · log(wᵢ + ε) Shannon entropy over normalized volume rendering weights wi. Higher entropy → frame is more uncertain → prioritized for keyframe selection.
02 · Reconstruction

Reconstructed mesh

Dense semantic reconstruction from Replica room_1. Drag to orbit · scroll to zoom · right-drag to pan.

loading model…
03 · Trajectory

Camera trajectory comparison

Estimated vs. ground-truth camera path. Drag to orbit · scroll to zoom. ━ estimated   ━ ground truth

Baseline (SNI-SLAM)
—
ATE RMSE (cm)
—
ATE Mean (cm)
Our method
Uncertainty-aware
—
ATE RMSE (cm)
—
ATE Mean (cm)
BASELINE
OURS
04 · Results

Quantitative evaluation

Evaluated on Replica dataset room_1 · NVIDIA Tesla V100 · Nstratified=32, Nimportance=8

BaselineSNI-SLAM (w/ GT seg)
ATE Mean0.7454 cm
ATE RMSE1.10 cm
Depth L11.21 cm
Completeness90.73%
FPS0.88
OursUncertainty-aware SLAM
ATE Mean0.5368 cm↓28%
ATE RMSE0.6833 cm↓38%
Depth L11.21 cm
Completeness90.44%
FPS0.68
Absolute Trajectory Error (ATE) — lower is better
BaselineATE Mean: 0.7454 cm
OursATE Mean: 0.5368 cm  ↓28%
Runtime — higher FPS is better
Baseline (SNI-SLAM)
0.88 FPS100%
+ Uncertainty est.
0.46 FPS52%
+ Vectorization
0.68 FPS77%
05 · Contributions

What we built

1
Uncertainty-aware keyframe selection — integrates Shannon entropy with covisibility to identify frames with the most informative rendering uncertainty for targeted network updates.
2
Convergence improvement under underfitting — high-uncertainty frames produce more informative gradient updates, helping the model prioritize geometrically ambiguous regions during online SLAM.
3
Fully vectorized batch entropy computation — reduces Nkf sequential decoder forward passes to a single batched pass, recovering 25% of the FPS overhead from naive uncertainty estimation.
4
Empirical validation on Replica — ATE Mean 0.7454→0.5368 cm (↓28%) and ATE RMSE 1.10→0.6833 cm (↓38%), with comparable reconstruction quality.