Research map / 3D vision / Depth and stereo
PORF: POSE RESIDUAL FIELD FOR ACCURATE NEURAL SURFACE RECONSTRUCTION
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 PoRF, an MLP that regresses per-frame pose residuals shared across the sequence, plus an epipolar geometry loss on COLMAP-exported correspondences, to refine camera poses for neural surface reconstruction.
Problem
Neural surface reconstruction is highly sensitive to camera pose noise, even with COLMAP or ARKit poses, and existing joint pose-NeRF optimisation methods fail to improve pose accuracy on challenging real-world captures.
Approach
Replace independent per-frame pose parameters with a shared MLP (Pose Residual Field) that takes frame index and initial pose and outputs a pose residual, and supervise poses with an epipolar (Sampson) loss on sparse SIFT matches exported from COLMAP, in addition to the NeuS colour loss.
Key results
- On DTU, rotation error reduced by 78% from COLMAP poses (0.65 to 0.14 degrees mean), and Voxurf reconstruction Chamfer distance drops from 3.48mm to 0.85mm.
- On MobileBrick, refining ARKit poses raises Voxurf reconstruction F1 from 69.18 to 75.67, above the provided GT pose (75.13 in the table, 75.14 in text).
- Integrated into Nerfstudio's Nerfacto, PoRF improves mean PSNR from 20.32 to 21.54 across 10 scenes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Refine COLMAP poses with PoRF and epipolar geometry loss | PoRF · DTU | rotation error (Deg.): 0.65 -> 0.14 | Helped |
| Refine COLMAP poses with PoRF and epipolar geometry loss | PoRF · DTU | Chamfer distance (mm): 3.48 -> 0.85 | Helped |
| Refine ARKit poses with PoRF and epipolar geometry loss | PoRF · MobileBrick, σ=2.5mm | F1: 69.18 -> 75.67 | Helped |
| Replace Nerfacto pose optimisation with PoRF MLP | Nerfacto · Nerfstudio, mean | PSNR: 20.32 -> 21.54 | Helped |
| Add PoRF to naive joint pose-parameter optimisation (ablation) | DTU pose refinement baseline · DTU | rotation error (Deg.): 0.38 -> 0.19 | Helped |
| Add epipolar geometry loss to naive joint pose-parameter optimisation (ablation) | DTU pose refinement baseline · DTU | rotation error (Deg.): 0.38 -> 0.20 | Helped |
| Replace SIFT correspondences with LoFTR correspondences (ablation) | Ours · DTU | rotation error (Deg.): 0.14 -> 0.13 | No clear effect |
| Replace SDF representation with NeRF representation (ablation) | Ours · DTU | rotation error (Deg.): 0.14 -> 0.14 | No clear effect |
| Remove the initial camera pose input (ablation) | PoRF · DTU Scan37 | performance | Hurt |
| Remove the frame index input (ablation) | PoRF · DTU Scan37 | performance | Hurt |
Surprising findings
- Reconstruction accuracy is highly sensitive to rotation error but only weakly to translation error, so translation gains are minor.
- BARF and L2G fail to improve COLMAP poses on DTU and L2G diverges in several scenes.
Limitations
- Training is slow (about 2.5 hours) because it builds on NeuS.
- Requires nearly dense views and a robust initial pose; not designed for sparse views unlike SPARF.
- Translation refinement gains are minor because reconstruction loss gives weak translation supervision.
Experimental setup
Datasets: DTU MobileBrick Nerfstudio dataset
Benchmarks: DTU MobileBrick Nerfstudio dataset
Metrics: rotation error (deg) translation error (mm) Chamfer distance (mm) F1 score PSNR SSIM LPIPS
Baselines: BARF L2G SPARF Nope-NeRF NeRF IDR VolSDF PointNeRF NeuS Voxurf TSDF-Fusion BNV-Fusion
Builds on: NeuS BARF Voxurf NeRF Nerfstudio
Closest papers by meaning
- NeuSurf: On-Surface Priors for Neural Surface Reconstruction from Sparse Input Views (AAAI 2024 · similarity 0.87)
- ProbeSDF: Light Field Probes For Neural Surface Reconstruction (CVPR 2025 · similarity 0.86)
- SC-NeuS: Consistent Neural Surface Reconstruction from Sparse and Noisy Views (AAAI 2024 · similarity 0.86)
- NeuRodin: A Two-stage Framework for High-Fidelity Neural Surface Reconstruction (NeurIPS 2024 · similarity 0.85)
- Neural Refinement for Absolute Pose Regression with Feature Synthesis (CVPR 2024 · similarity 0.85)
- NRDF: Neural Riemannian Distance Fields for Learning Articulated Pose Priors (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.