Research map / 3D vision / NeRF and radiance fields
Neural Visibility Field for Uncertainty-Driven Active Mapping
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 Neural Visibility Field (NVF), a NeRF uncertainty method that uses a Bayesian network over volume rendering and a learned visibility head so unobserved regions get high uncertainty, applied to next-best-view active mapping.
Problem
Quantifying NeRF uncertainty for active mapping in a principled way, especially for regions never visible in training views, where existing proxies (density, position-based color variance) fail to account for visibility.
Approach
Model volume rendering as a Bayesian network so that position-based color and occlusion uncertainty composite into a ray-level Gaussian mixture; add a learned visibility field so invisible points receive a prior color distribution, and pick next views by maximizing GMM entropy.
Key results
- On the Room scene NVF reaches 22.83 PSNR and 0.943 SSIM versus 15.19 PSNR for the best baseline (AIR).
- On the Hubble scene NVF reaches 27.99 PSNR and visual coverage 0.681 versus 24.63 PSNR and 0.586 for AIR.
- Ablation on active mapping shows removing the visibility term drops PSNR from 24.42 to 21.11 and visual coverage from 0.546 to 0.382.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NVF instead of ActiveRMAP | Active mapping with Instant-NGP · NeRF Assets (Avg.) | PSNR: 20.03 -> 23.90 | Helped |
| Use NVF instead of NeurAR | Active mapping with Instant-NGP · Hubble | PSNR: 25.19 -> 27.99 | Helped |
| Use NVF instead of AIR | Active mapping with Instant-NGP · Room | PSNR: 15.19 -> 22.83 | Helped |
| Add the visibility factor (ablation) | NVF active mapping · All scenes (average) | PSNR: 21.11 -> 24.42 | Helped |
| Add spatial color variance estimation (ablation) | NVF active mapping · All scenes (average) | PSNR: 23.77 -> 24.42 | Helped |
| Add the ray-correlation correction (ablation) | NVF active mapping · All scenes (average) | PSNR: 20.32 -> 24.42 | Helped |
| Use a tighter entropy upper bound (ablation) | NVF active mapping · All scenes (average) | PSNR: 22.54 -> 24.42 | Helped |
Surprising findings
- Several baselines (ActiveNeRF, NeurAR) assign similar or even lower uncertainty to unobserved regions than observed ones in complex scenes.
- Removing the NeRF-predicted color variance has a comparatively small effect, while removing visibility causes the largest drop.
Limitations
- Active mapping pipeline does not account for trajectory constraints or path cost.
- Requires assumed coarse bounding box of the region of interest and a fixed candidate sampling without constraints.
Experimental setup
Datasets: NeRF synthetic dataset Hubble Space Telescope scene (Blender) custom Room scene (Blender)
Benchmarks: NeRF synthetic assets Hubble scene Room scene
Metrics: PSNR SSIM LPIPS RGB loss Accuracy Completion Completion Ratio visual coverage
Baselines: WD ActiveRMAP AIR ActiveNeRF NeurAR Random
Builds on: NeRF Instant-NGP ActiveNeRF ActiveRMAP Active Implicit Reconstruction
Closest papers by meaning
- Bayes' Rays: Uncertainty Quantification for Neural Radiance Fields (CVPR 2024 · similarity 0.89)
- ProvNeRF: Modeling per Point Provenance in NeRFs as a Stochastic Field (NeurIPS 2024 · similarity 0.88)
- NARUTO: Neural Active Reconstruction from Uncertain Target Observations (CVPR 2024 · similarity 0.86)
- Learning Robust Generalizable Radiance Field with Visibility and Feature Augmented Point Representation (ICLR 2024 · similarity 0.86)
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.86)
- NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild (CVPR 2024 · similarity 0.85)
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.