Research map / 3D vision / Depth and stereo
NARUTO: Neural Active Reconstruction from Uncertain Target Observations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NARUTO, a neural active reconstruction system that uses a hash-grid hybrid representation with a learned uncertainty volume to drive goal search and RRT path planning for 6DoF exploration, plus an active ray sampling strategy for neural SLAM and mapping.
Problem
Active reconstruction of indoor scenes with neural representations is slow and restricts agent motion to low-DoF settings (fixed points, hemispheres, or 2D planes); the goal is high-completeness, high-fidelity surface reconstruction with unrestricted 6DoF movement.
Approach
Combine a Co-SLAM-style hybrid hash-grid SDF/color map with an explicit learnable uncertainty voxel grid updated through uncertainty-aware depth rendering; aggregate uncertainty over a multi-level goal space for goal selection, plan with an efficient RRT on the SDF volume, and replace random ray sampling with top-uncertainty rays.
Key results
- On MP3D, NARUTO reaches 90.18% completion ratio versus 73.15% for ANM, with completion 3.00 cm vs 9.11 cm.
- On Replica, the full method achieves 97.20% completion ratio, versus 92.74% for Co-SLAM and 96.89% for the variant without active ray sampling.
- Active ray sampling added to Co-SLAM reduces variance of accuracy and completion across trials (e.g., completion-ratio variance 72.90e-2 to 11.77e-2 in the SLAM setting).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NARUTO instead of ANM | MP3D active reconstruction · MP3D | Completion ratio (%): 73.15 -> 90.18 | Helped |
| Use NARUTO instead of ANM | MP3D active reconstruction · MP3D | Completion (cm): 9.11 -> 3.00 | Helped |
| Add Active Ray Sampling (ablation) | NARUTO active neural mapping · Replica | Completion ratio (%): 96.89 -> 97.20 | Helped |
| Add Active Ray Sampling (ablation) | NARUTO active neural mapping · Replica | Completion (cm): 1.76 -> 1.66 | Helped |
| Replace the explicit uncertainty grid with an implicit uncertainty network (ablation) | NARUTO active neural mapping · Replica | Completion ratio (%): 97.20 -> 94.62 | Hurt |
| Replace the explicit uncertainty grid with an implicit uncertainty network (ablation) | NARUTO active neural mapping · Replica | Completion (cm): 1.66 -> 2.05 | Hurt |
| Add Active Ray Sampling to Co-SLAM | Co-SLAM neural SLAM · Replica | Completion ratio variance: 72.90 -> 11.77 | Helped |
| Add Active Ray Sampling to Co-SLAM | Co-SLAM neural SLAM · Replica | Completion ratio (%): 92.74 -> 92.70 | No clear effect |
Surprising findings
- The implicit uncertainty MLP produces fluctuating and unobserved regions with random uncertainty, which motivated the explicit uncertainty grid.
- MP3D ground-truth meshes are incomplete, so neural extrapolation penalizes the Accuracy metric even when reconstruction is better.
Limitations
- Assumes known localization and perfect action execution.
- Agent motion constraints are not modeled in the current system.
- Uses a single-resolution uncertainty grid.
Experimental setup
Datasets: Replica Matterport3D
Benchmarks: Replica Matterport3D
Metrics: Accuracy (cm) Completion (cm) Completion ratio (%) MAD (cm)
Baselines: ANM iMAP NICE-SLAM Co-SLAM FBE UPEN OccAnt ActiveNeRF NeuRAR
Builds on: Co-SLAM Active Neural Mapping RRT-Connect NeRF Kendall and Gal depth uncertainty
Closest papers by meaning
- Neural Visibility Field for Uncertainty-Driven Active Mapping (CVPR 2024 · similarity 0.86)
- Understanding while Exploring: Semantics-driven Active Mapping (NeurIPS 2025 · similarity 0.83)
- SC-NeuS: Consistent Neural Surface Reconstruction from Sparse and Noisy Views (AAAI 2024 · similarity 0.83)
- Swift-Mapping: Online Neural Implicit Dense Mapping in Urban Scenes (AAAI 2024 · similarity 0.83)
- AE-NeRF: Augmenting Event-Based Neural Radiance Fields for Non-ideal Conditions and Larger Scenes (AAAI 2025 · similarity 0.83)
- NextBestPath: Efficient 3D Mapping of Unseen Environments (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.