Research map / 3D vision / NeRF and radiance fields
NeRF Director: Revisiting View Selection in Neural Volume Rendering
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 NeRF Director, a framework for training-view selection in NeRF (training-free farthest view sampling and information-gain sampling with relaxation) plus a more uniform test-view split, showing view selection changes NeRF rankings and speeds convergence.
Problem
Training-view selection for neural volume rendering is understudied, and test-view placement can bias evaluation and reshuffle SOTA rankings of NeRF methods.
Approach
A training-free farthest view sampling that combines camera-center distance with SfM-based photogrammetric view similarity to pick diverse training views, alongside an error-guided information gain sampler with a Lloyd-based relaxation step to avoid clustering, and a uniform sphere-based test split.
Key results
- Choice of view selection scheme changes PSNR by up to 1.9 dB on the benchmarks.
- FVS reaches the converged quality of random sampling with a fixed view budget up to 4x faster.
- Uncertainty-based ActiveNeRF and Density-aware NeRF Ensembles underperform random sampling on NeRF Synthetic with InstantNGP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Select training views with FVS instead of random sampling | InstantNGP · NeRF Synthetic, synthetic setting | PSNR: 70 vs 150 views to reach 33 dB | Helped |
| Use FVS rather than random sampling under the same view budget | NeRF · NeRF Synthetic and TanksAndTemples | speedup: up to 4× speedup | Helped |
| Use error rather than uncertainty as the information signal, with the vMF sampler and relaxation (ablation) | IGS · TanksAndTemples | PSNR | Helped |
| Use a sampler that considers nearby error measurements, with relaxation, rather than the compared alternative sampling strategy (ablation) | IGS · TanksAndTemples | PSNR | Helped |
| Use Euclidean spatial distance rather than great-circle distance alone (ablation) | FVS · TanksAndTemples | PSNR | Helped |
| Add photogrammetric distance to camera-center distance (ablation) | FVS · TanksAndTemples, Playground | PSNR | Helped |
| Use uncertainty-guided ActiveNeRF view selection instead of random sampling | InstantNGP · NeRF Synthetic | PSNR | Hurt |
| Use uncertainty-guided Density-aware NeRF Ensembles instead of random sampling | InstantNGP · NeRF Synthetic | PSNR | Hurt |
| Use Density-aware NeRF Ensembles instead of random sampling | InstantNGP · TanksAndTemples, more than 60 views | PSNR | Helped |
| Use FVS instead of random sampling | InstantNGP · TanksAndTemples, more than 30 views | PSNR | Helped |
| Apply relaxation after proposing new views (ablation) | IGS · TanksAndTemples | PSNR | Helped |
| Use greedy information-gain selection without relaxation (ablation) | IGS · NeRF view selection | PSNR | Hurt |
Surprising findings
- Uncertainty-guided view selection (ActiveNeRF, Density-aware NeRF Ensembles) performs worse than random sampling on NeRF Synthetic.
- A simple 90-degree rotation of test views can reverse SOTA method rankings on NeRF Synthetic.
Limitations
- Methods assume same-resolution cameras and an object-centric setting.
- FVS may be limited in complex indoor scenes where camera-center distance fails to capture view similarity.
- IGS depends heavily on the accuracy of error or uncertainty information.
Experimental setup
Datasets: NeRF Synthetic TanksAndTemples
Benchmarks: NeRF Synthetic TanksAndTemples
Metrics: PSNR SSIM
Baselines: random sampling ActiveNeRF Density-aware NeRF Ensembles NeRF Mip-NeRF Plenoxels InstantNGP
Builds on: Farthest point sampling Active-NeRF Density-aware NeRF Ensembles Lloyd-Max algorithm
Closest papers by meaning
- NeRF-VPT: Learning Novel View Representations with Neural Radiance Fields via View Prompt Tuning (AAAI 2024 · similarity 0.88)
- NerfBaselines: Consistent and Reproducible Evaluation of Novel View Synthesis Methods (NeurIPS 2025 · similarity 0.88)
- Rad-NeRF: Ray-decoupled Training of Neural Radiance Field (NeurIPS 2024 · similarity 0.88)
- GO-N3RDet: Geometry Optimized NeRF-enhanced 3D Object Detector (CVPR 2025 · similarity 0.87)
- NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild (CVPR 2024 · similarity 0.87)
- CodecNeRF: Toward Fast Encoding and Decoding, Compact, and High-quality Novel-view Synthesis (AAAI 2025 · similarity 0.87)
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.