Research map / Point clouds / Point cloud classification
Noisy One-point Homographies are Surprisingly Good
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 a one-point minimal solver for calibrated Euclidean homography estimation that combines SIFT scale/orientation as relative depth and affine constraints, used inside GC-RANSAC to beat 2-AC and 4-point on HEB and HPatches with lower runtime.
Problem
Two-view homography estimation with multiple planes yields very low inlier ratios per plane, requiring many RANSAC iterations; reducing the minimal sample size is key.
Approach
Derive a one-point solver by reusing the same SIFT scale and orientation in two approximately-dependent constraints (relative depth from scale, and first-order affine constraints from scale/orientation, plus a line constraint), accepting very noisy homographies that a robust local optimizer (GC-RANSAC) can refine.
Key results
- On HEB, the 1-SIFT solver in GC-RANSAC reaches mAA@10° of 29.4, matching 2-AC (29.1) while running in 10.3 ms versus 15.3 ms (about 30% faster).
- On HPatches with SuperPoint+LightGlue keypoints, the 4-point solver reaches 79.3 full-set AUC at 5px versus 76.8 for 1-SIFT with Self-Sca-Ori scale/orientation, so the 4-point solver is stronger there.
- The 1-SIFT solver reaches 14.7 mAA@5° on HEB, compared with 13.4 for the 2-SIFT baseline and 11.3 for the 4-point solver.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the one-point 1-SIFT solver instead of the 2-AC solver | GC-RANSAC on HEB · HEB | Runtime (ms): 15.3 -> 10.3 | Helped |
| Use the one-point 1-SIFT solver instead of the 2-AC solver | GC-RANSAC on HEB · HEB | mAA@10°: 29.1 -> 29.4 | Helped |
| Use SIFT scale and orientation rather than random scale and orientation (ablation) | 1-SIFT on HPatches viewpoint pairs · HPatches viewpoint | Corner-warp AUC@5px: 13.5 -> 63.5 | Helped |
| Use SIFT scale and orientation rather than random scale and orientation (ablation) | 1-SIFT on HPatches illumination pairs · HPatches illumination | Corner-warp AUC@5px: 8.13 -> 72.0 | Helped |
| Replace SIFT scale and orientation with fixed scale 1.0 and SIFT orientation (ablation) | 1-SIFT on HPatches viewpoint pairs · HPatches viewpoint | Corner-warp AUC@5px: 63.5 -> 56.1 | Hurt |
| Replace SIFT orientation with random orientation (ablation) | 1-SIFT on HPatches viewpoint pairs · HPatches viewpoint | Corner-warp AUC@5px: 63.5 -> 13.1 | Hurt |
| Replace SIFT scale with random scale (ablation) | 1-SIFT on HPatches viewpoint pairs · HPatches viewpoint | Corner-warp AUC@5px: 63.5 -> 61.3 | Hurt |
| Use Self-Sca-Ori instead of SIFT scale and orientation (ablation) | 1-SIFT on HPatches · HPatches full | Corner-warp AUC@5px: 67.3 -> 66.4 | Hurt |
| Use H-Similarity instead of the 1-SIFT solver (ablation) | homography estimation on HPatches full · HPatches full | Corner-warp AUC@5px: 67.3 -> 66.9 | No clear effect |
| Use HSolo instead of the 1-SIFT solver (ablation) | homography estimation on HPatches full · HPatches full | Corner-warp AUC@5px: 67.3 -> 51.0 | Hurt |
| Use 4-point instead of 1-SIFT | homography estimation on HPatches full with SIFT features · HPatches full | Corner-warp AUC@5px: 67.3 -> 71.3 | Helped |
| Refine local affine shape with AffNet (ablation) | 1-SIFT on HEB · HEB | mAA(R)@10°: 33.4 -> 37.3 | Helped |
Surprising findings
- Solvers that produce very noisy homographies from a single correspondence are still sufficient, because GC-RANSAC local optimization converges to good models from them.
- An intentionally incorrect line constraint (treating line normals as points) has only a very minor impact on results.
Limitations
- Requires known intrinsic calibration (focal length and principal point), though coarse estimates suffice.
- Requires each keypoint to have scale and orientation, which excludes many learned detectors that only give positions.
- Estimates from self-estimated scale/orientation (Self-Sca-Ori) are coarse.
Experimental setup
Datasets: HEB (Homography Estimation Benchmark) HPatches
Benchmarks: HEB HPatches
Metrics: mAA@5° mAA@10° mAA@2m mAA@5m AUC corner-warp error runtime (ms) number of inliers
Baselines: 4-point solver 3-SIFT 2-SIFT 2-AC HSolo H-Similarity
Builds on: GC-RANSAC affine correspondence homography solver (2-AC) SIFT 4-point homography solver Self-Sca-Ori
Closest papers by meaning
- Mask-Homo: Pseudo Plane Mask-Guided Unsupervised Multi-Homography Estimation (AAAI 2024 · similarity 0.88)
- Adapting Dense Matching for Homography Estimation with Grid-based Acceleration (CVPR 2025 · similarity 0.88)
- MCNet: Rethinking the Core Ingredients for Accurate and Efficient Homography Estimation (CVPR 2024 · similarity 0.87)
- HomoMatcher: Achieving Dense Feature Matching with Semi-Dense Efficiency by Homography Estimation (AAAI 2025 · similarity 0.87)
- Unsupervised Homography Estimation on Multimodal Image Pair via Alternating Optimization (NeurIPS 2024 · similarity 0.87)
- Three-view Focal Length Recovery From Homographies (CVPR 2025 · 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.