Research map / 3D vision / Depth and stereo
Structure-from-Motion with a Non-Parametric Camera Model
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 GenSfM, an incremental structure-from-motion pipeline that self-calibrates a spline-based non-parametric radial camera model adaptively, handling pinhole through fisheye and catadioptric cameras.
Problem
Structure-from-motion with parametric camera models fails or is inaccurate for highly distorted, fisheye, or catadioptric cameras whose intrinsics are unknown.
Approach
Model the radial angle-to-radius map with an adaptive cubic spline that is calibrated only over image regions with sufficient constraints, and triangulate with mixed 2D and distortion-invariant radial (1D) constraints, bootstrapped from a radial quadrifocal solver and implicit distortion pose estimation.
Key results
- On BabelCalib, GenSfM reaches reprojection errors comparable to BabelCalib (e.g., Kalibr 0.25 vs 0.21) and much lower than implicit model [18] on most sets.
- On ETH3D DSLR pose AUC@5°, GenSfM averages 71.8 vs 61.5 for COLMAP thin prism and 50.0 for COLMAP radial, the highest average.
- On artificially distorted ETH3D, GenSfM keeps pose AUC@5° at 43.2 average while COLMAP variants average 2.4 or lower and RadialSfM+[18] 14.5 with the other baselines much lower.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GenSfM instead of COLMAP (thin prism) | SfM on ETH3D · ETH3D | AUC @ 5° camera pose: 61.5 -> 71.8 | Helped |
| Use GenSfM instead of RadialSfM + implicit model | SfM on distorted ETH3D · Distorted ETH3D | AUC @ 5° camera pose: 14.5 -> 43.2 | Helped |
| Use GenSfM instead of COLMAP (radial) | SfM on ETH3D · ETH3D | AUC completeness: 21.5 -> 24.5 | Helped |
| Use GenSfM instead of COLMAP (radial) | SfM on ETH3D · ETH3D | AUC accuracy: 72.6 -> 80.9 | Helped |
| Use GenSfM instead of COLMAP (radial) | SfM on ETH3D · ETH3D | AUC F1: 30.7 -> 35.1 | Helped |
| Vary the number of calibration control points (ablation) | GenSfM calibration on BabelCalib · BabelCalib | projection error | Mixed |
| Use 10 calibration control points instead of other tested counts (ablation) | GenSfM calibration · BabelCalib | projection error | Helped |
Surprising findings
- Estimating principal point is not necessary: reprojection error stays low with principal points fixed at image center, even when true principal points deviate by tens of pixels.
Limitations
- Single calibrated segment only; multiple disjoint calibrated regions are not used.
- Bundle adjustment fixes theta values of control points and only optimizes radii, so monotonicity is not guaranteed and needs postprocessing.
- Some scenes (meadow, pipes, terrains) fail or have poor results, with initialization failures for some baselines.
Experimental setup
Datasets: BabelCalib ETH3D DSLR ETH3D distorted (synthetic radial distortion) fisheye datasets from RadialSfM catadioptric images
Benchmarks: BabelCalib ETH3D
Metrics: reprojection RMS error ratio of calibrated points AUC@5 pose error point cloud completeness/accuracy/F1 AUC
Baselines: COLMAP (pinhole) COLMAP (radial) COLMAP (thin prism) RadialSfM + implicit model BabelCalib
Builds on: RadialSfM implicit distortion model COLMAP radial quadrifocal tensor solver
Closest papers by meaning
- Non-Rigid Structure-from-Motion: Temporally-Smooth Procrustean Alignment and Spatially-Variant Deformation Modeling (CVPR 2024 · similarity 0.87)
- VGGSfM: Visual Geometry Grounded Deep Structure From Motion (CVPR 2024 · similarity 0.86)
- Deep Non-Rigid Structure-from-Motion Revisited: Canonicalization and Sequence Modeling (AAAI 2025 · similarity 0.86)
- Leveraging Camera Triplets for Efficient and Accurate Structure-from-Motion (CVPR 2024 · similarity 0.84)
- AnyMap: Learning a General Camera Model for Structure-from-Motion with Unknown Distortion in Dynamic Scenes (CVPR 2025 · similarity 0.84)
- Light3R-SfM: Towards Feed-forward Structure-from-Motion (CVPR 2025 · similarity 0.84)
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.