Research map / 3D vision / NeRF and radiance fields
CF-NeRF: Camera Parameter Free Neural Radiance Fields with Incremental Learning
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 CF-NeRF, an end-to-end NeRF that jointly estimates camera parameters and scene representation by incrementally registering images one by one, inspired by incremental SfM, and evaluates on the NeRFBuster rotation dataset.
Problem
NeRF methods that jointly learn camera parameters (NeRFmm, BARF, L2G-NeRF, etc.) work only for forward-facing scenes with small motion and fail under complex rotational camera trajectories, while standard NeRF requires COLMAP poses.
Approach
Recovers camera poses one image at a time, mirroring incremental SfM: initialize on a few images, implicitly localize each new image by optimizing only its pose against a fixed NeRF, then refine with partial and periodic global implicit optimization within a coarse-to-fine Gaussian pyramid.
Key results
- On NeRFBuster, CF-NeRF reduces average rotation error ΔR to roughly 10 degrees versus around 100 degrees for baselines, per the paper's text.
- CF-NeRF achieves PSNR of 26.94 on aloe and 32.37 on pipe-scene (pikachu) on NeRFBuster, the best among compared methods on all 12 scenes for PSNR and LPIPS.
- On LLFF, CF-NeRF is competitive with NeRFmm and BARF on forward-facing scenes, and much better on the two rotation scenes (vasedeck, pinecone).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase iterations ξ from 600 to 900 (ablation) | CF-NeRF · NeRFBuster | ΔR: 17.8029 -> 14.8730 | Helped |
| Reduce global optimization interval Nglob from 10 to 5 (ablation) | CF-NeRF · NeRFBuster | ΔR: 14.8730 -> 12.1226 | Helped |
| Use coarse-to-fine optimization rather than coarse stage only (ablation) | CF-NeRF · NeRFBuster | ΔR: 12.4862 -> 12.1226 | Helped |
| Use CF-NeRF instead of L2G-NeRF | NeRFBuster aloe · NeRFBuster aloe | ΔR: 124.2368 -> 12.1226 | Helped |
| Use CF-NeRF instead of BARF | NeRFBuster aloe · NeRFBuster aloe | ΔT: 9.6196 -> 3.3788 | Helped |
| Use CF-NeRF instead of SiRENmm | NeRFBuster aloe · NeRFBuster aloe | PSNR: 22.7462 -> 26.9367 | Helped |
| Use CF-NeRF instead of BARF | NeRFBuster aloe · NeRFBuster aloe | LPIPS: 0.3328 -> 0.1939 | Helped |
Surprising findings
- Baselines often achieve reasonable novel-view PSNR despite very poor camera pose estimates, attributed to NeRF's overfitting ability.
Limitations
- Remaining gap between CF-NeRF camera estimates and COLMAP.
- Sensitive to iteration count and global optimization frequency; training is time-consuming since each image needs many iterations.
- Mixed results on LLFF forward-facing scenes compared with NeRFmm and BARF.
Experimental setup
Datasets: NeRFBuster LLFF Real360
Benchmarks: NeRFBuster LLFF Real360
Metrics: rotation error (ΔR) translation error (ΔT) PSNR LPIPS
Baselines: NeRFmm SiRENmm BARF GARF L2G-NeRF
Builds on: NeRF NeRFmm SiRENmm BARF iNeRF
Closest papers by meaning
- Joint Optimization of Neural Radiance Fields and Continuous Camera Motion from a Monocular Video (CVPR 2025 · similarity 0.91)
- PyNeRF: Pyramidal Neural Radiance Fields (NeurIPS 2023 · similarity 0.91)
- D-NeRF: Neural Radiance Fields for Dynamic Scenes (CVPR 2021 · similarity 0.91)
- NoPe-NeRF: Optimising Neural Radiance Field With No Pose Prior (CVPR 2023 · similarity 0.91)
- NeRF-VPT: Learning Novel View Representations with Neural Radiance Fields via View Prompt Tuning (AAAI 2024 · similarity 0.91)
- Mask-Based Modeling for Neural Radiance Fields (ICLR 2024 · similarity 0.90)
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.