Research map / 3D vision / Depth and stereo
Leveraging Camera Triplets for Efficient and Accurate Structure-from-Motion
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 camera-triplet edge scoring method with a provably thresholding-based edge selection that simultaneously sparsifies SfM viewgraphs and removes false edges from repeated structures.
Problem
Structure-from-Motion viewgraphs contain many redundant edges, which slow reconstruction, and false edges from incorrect retrieval and repeated structures, which cause ghosting and superimposed reconstructions. Existing methods handle these separately and often need intermediate reconstructions or m
Approach
Score each edge by the ratio of its epipolar inlier count to the maximum inlier count within each triplet it belongs to, average over triplets, and show that the regularized selection problem is solved exactly by thresholding, with a single adaptive threshold derived from max node degree.
Key results
- On generic datasets, the sparsified graph GF(0.6) reduces COLMAP reconstruction time by roughly 50% to 80% while keeping most cameras and 3D points and removing ghost artifacts.
- On ambiguous datasets, the method disambiguates all large and medium-scale datasets except Cup, recovering more cameras and points than Doppelgangers in most cases.
- The method takes under 1% of total reconstruction time in most cases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Sparsify the viewgraph using camera-triplet edge scores | COLMAP SfM on generic datasets · Generic datasets | reconstruction time: 44 -> 28; 372 -> 144; 6875 -> 3698 | Helped |
| Sparsify the viewgraph using camera-triplet edge scores | COLMAP SfM on generic datasets · Generic datasets | mean reprojection error: 0.66 -> 0.62; 0.76 -> 0.61 | Helped |
| Sparsify the viewgraph using camera-triplet edge scores | COLMAP SfM on generic datasets · Notre Dame | cameras reconstructed: 1407 -> 1093 | Hurt |
| Remove ambiguous and redundant edges using camera-triplet scores | COLMAP SfM · Louvre | 3D points reconstructed: 108 -> 113 | Helped |
| Remove ambiguous and redundant edges using camera-triplet scores | COLMAP SfM · Notre Dame (ambiguous) | reconstruction time: 1160 -> 1059 | Helped |
| Apply the method to Cup | COLMAP SfM · Cup | disambiguation | Hurt |
| Increase minimum edge score from 0.6 to 0.7 (ablation) | Ours on Alamo · Alamo | reconstruction time: 28 -> 25 | Helped |
| Increase minimum edge score from 0.6 to 0.7 (ablation) | Ours on Gendarmenmarkt · Gendarmenmarkt | reconstruction time: 144 -> 103 | Helped |
| Increase minimum edge score from 0.6 to 0.7 (ablation) | Ours on Union Square · Union Square | reconstruction time: 80 -> 86 | Hurt |
| Increase minimum edge score from 0.6 to 0.7 (ablation) | Ours on generic datasets · Generic datasets | cameras reconstructed | Hurt |
| Remove edges not contributing to the largest connected component of the triplet graph (ablation) | Ours · Generic datasets | reconstruction quality | No clear effect |
Surprising findings
- Separating joints and keeping only the largest triplet-graph component has minimal effect on reconstructions.
- Doppelgangers could not disambiguate some datasets for any threshold, while the proposed method disambiguated all but Cup.
Limitations
- Some small-scale sequential datasets are over-split after sparsification due to low edge redundancy.
- Edges not part of any triplet are removed, and joints are separated, which may discard valid edges.
- Relies on COLMAP for viewgraph and reconstruction in experiments.
Experimental setup
Datasets: Generic datasets from Heinly et al. (Alamo, Gendarmenmarkt, Notre Dame, etc.) Dubrovnik Rome Quad Ambiguous datasets from Doppelgangers Louvre Notre Dame (Wilson et al.) Sacre Coeur Seville Alexander Nevsky Cathedral Books/Cereal/Cup (Yan et al.)
Benchmarks: Generic SfM datasets (Table 2) Ambiguous SfM datasets (Table 3)
Metrics: number of cameras reconstructed number of 3D points reconstructed mean reprojection error reconstruction time
Baselines: Doppelgangers COLMAP original viewgraph
Builds on: COLMAP Doppelgangers
Closest papers by meaning
- Light3R-SfM: Towards Feed-forward Structure-from-Motion (CVPR 2025 · similarity 0.87)
- Detector-Free Structure from Motion (CVPR 2024 · similarity 0.86)
- Learning to Filter Outlier Edges in Global SfM (CVPR 2025 · similarity 0.86)
- Graph-Guided Scene Reconstruction from Images with 3D Gaussian Splatting (ICLR 2025 · similarity 0.85)
- SPARS3R: Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction (CVPR 2025 · similarity 0.85)
- VGGSfM: Visual Geometry Grounded Deep Structure From Motion (CVPR 2024 · 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.