Research map / 3D vision / Human pose estimation
Absolute Pose from One or Two Scaled and Oriented Features
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.
Derives minimal absolute pose solvers from one or two scaled and oriented feature correspondences (with surface normal and depth known), P2ORI for the general case and UP1SIFT for the gravity-aware case, and shows gains in RANSAC localization on Cambridge Landmarks and Aachen Day-Night.
Problem
Absolute camera pose estimation normally needs three 2D-3D point correspondences (P3P), so RANSAC sample sizes are large in high-outlier settings; the paper asks whether scale and orientation from keypoints can reduce the minimal sample size.
Approach
Derives constraints linking absolute pose to scaled and oriented features by combining the affine-correspondence pose equations of P1AC with the affine-to-scale/orientation relations of Barath and Kukelova, yielding a two-point general solver (P2ORI) and a one-point gravity-aware solver (UP1SIFT).
Key results
- On Cambridge Landmarks with SP+LG+SSO features, P2ORI reaches 80.1% average recall at 0.05m/1deg versus 76.5% for P1AC and 73.9% for P3P (GC-RANSAC).
- On Aachen Night with SP+LG+SSO, P2ORI achieves 71.7 / 86.4 / 99.0 recall at 0.25m/2deg, 0.5m/5deg, 5m/10deg, the best among compared methods.
- Solver timings: P2ORI 2.62 us, P1AC 1.92 us, P3P 0.42 us, UP1SIFT 1.05 us, UP2P 0.23 us; all solvers run under 3 us.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use P2ORI instead of P3P | GC-RANSAC localization · Cambridge Landmarks weighted avg. | recall at 0.05m/1° (%): 39.5 -> 54.1 | Helped |
| Use UP1SIFT instead of P3P | GC-RANSAC localization · Cambridge Landmarks weighted avg. | recall at 0.1m/1° (%): 49.0 -> 69.9 | Helped |
| Use P2ORI instead of P3P | GC-RANSAC with RootSIFT · Aachen Day-Night Day | recall at 0.25m/2° (%): 55.6 -> 60.9 | Helped |
| Use UP1SIFT instead of P3P | GC-RANSAC with RootSIFT · Aachen Day-Night Night | recall at 0.25m/2° (%): 46.1 -> 47.6 | Helped |
| Use P2ORI instead of P3P | GC-RANSAC with SuperPoint + LightGlue + SelfScaleOri · Aachen Day-Night Day | recall at 0.25m/2° (%): 60.6 -> 68.2 | Helped |
| Use P2ORI instead of P1AC | GC-RANSAC with SuperPoint + LightGlue + SelfScaleOri · Aachen Day-Night Night | recall at 0.25m/2° (%): 69.6 -> 71.7 | Helped |
| Use P2ORI instead of P1AC | GC-RANSAC localization · Cambridge Landmarks weighted avg. | position error (cm): 25 -> 14 | Helped |
| Use UP1SIFT instead of P3P | GC-RANSAC localization · Cambridge Landmarks weighted avg. | position error (cm): 120 -> 4 | Helped |
Surprising findings
- P1AC works surprisingly well even though it is given approximated affine frames built from SIFT scale and orientation rather than true affine correspondences.
Limitations
- Solvers assume known depth and surface normal at the reference point, and are sensitive to orientation and log-scale noise.
- Gravity-aware UP1SIFT relies on a gravity direction (ground truth in Cambridge, assumed downward in Aachen).
Experimental setup
Datasets: Cambridge Landmarks Aachen Day-Night v1.1
Benchmarks: Cambridge Landmarks Aachen Day-Night v1.1
Metrics: median position error median rotation error recall at 0.05m/1deg recall at 0.1m/1deg recall at 0.2m/1deg recall at 0.25m/2deg recall at 0.5m/5deg recall at 5.0m/10deg runtime numerical stability
Baselines: P3P UP2P P1AC GC-RANSAC
Builds on: P1AC (Ventura et al. 2019) Barath and Kukelova affine-to-scale-orientation constraints UP2P (Kukelova et al. 2016) GC-RANSAC Cayley parameterization
Closest papers by meaning
- Efficient Solution of Point-Line Absolute Pose (CVPR 2024 · similarity 0.84)
- From Correspondences to Pose: Non-minimal Certifiably Optimal Relative Pose without Disambiguation (CVPR 2024 · similarity 0.84)
- FAR: Flexible Accurate and Robust 6DoF Relative Camera Pose Estimation (CVPR 2024 · similarity 0.84)
- Orient Anything V2: Unifying Orientation and Rotation Understanding (NeurIPS 2025 · similarity 0.83)
- Scene-agnostic Pose Regression for Visual Localization (CVPR 2025 · similarity 0.83)
- Practical Solutions to the Relative Pose of Three Calibrated Cameras (CVPR 2025 · similarity 0.83)
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.