Research map / Graph learning / Graph algorithms and clustering
Learning to Filter Outlier Edges in Global SfM
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.
Filters outlier relative-translation edges in global SfM by classifying nodes of a line graph with a small Transformer-based GNN, using graph clustering to scale to large pose graphs.
Problem
Outlier relative translation edges in the pose graph degrade translation averaging in global Structure-from-Motion, and existing filters (1DSfM, MSAC) give limited gains.
Approach
Reformulate edge filtering in the pose graph as vertex classification on its line graph (edges become nodes, cameras become edges), classify with a TransformerConv GNN using rotations and DINOv2 image embeddings, and apply clustering to bound line-graph memory.
Key results
- On 1DSfM with Theia, the proposed filter reduces average mean position error from 154.29 m (no filter) and 154.21 m (1DSfM) to 26.97 m, and Ours+Triangle reaches 21.15 m.
- On 1DSfM with GLOMAP, adding the proposed filter reduces median position error from 9.56 to 8.59 and raises Recall@10m from 59.39 to 61.68.
- On ScanNet, the proposed filter improves Recall@1m to 79.01% versus 77.57% without filtering and 72.05% with 1DSfM; combined with Triangle it reaches 85.71%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply the learned Ours filter | Theia · 1DSfM | mean position error: 154.21 -> 26.97 | Helped |
| add the Triangle filter to Ours | Theia · 1DSfM | Position Recall@10m: 40.78 -> 44.42 | Helped |
| apply the learned Ours filter | Theia · PhotoTourism | mean position error: 2.37 -> 2.33 | Helped |
| add the Triangle filter to Ours | Theia · PhotoTourism | Position Recall@5: 89.72 -> 88.86 | Hurt |
| apply the learned Ours filter | Theia · ScanNet | mean position error: 0.78 -> 0.73 | Helped |
| add the Triangle filter to Ours | Theia · ScanNet | Position Recall@1m: 79.01 -> 85.71 | Helped |
| add Ours filtering to default GLOMAP | GLOMAP · 1DSfM | median position error: 9.56 -> 8.59 | Helped |
| remove image features (ablation) | proposed filter · T. of London, M. Rushmore, E. Island | mean position error: 22.09 -> 37.78 | Hurt |
| use two GNN layers instead of three (ablation) | proposed filter · T. of London, M. Rushmore, E. Island | Recall@10: 51.39 -> 47.33 | Hurt |
| train with eight clusters instead of three (ablation) | proposed filter · T. of London, M. Rushmore, E. Island | mean position error: 22.09 -> 22.49 | Hurt |
| remove graph clustering at inference (ablation) | proposed filter · T. of London, M. Rushmore, E. Island | Recall@10: 51.39 -> 44.07 | Hurt |
Surprising findings
- Standalone 1DSfM slightly decreases accuracy on ScanNet relative to no filter, while the proposed filter improves all metrics there.
- The proposed filter combined with the Triangle filter retains fewer cameras (1263) than the Oracle filter (1341) yet achieves higher recall.
Limitations
- Slower than 1DSfM filtering (minutes vs. seconds).
- Oracle threshold (20 deg Theia, 10 deg GLOMAP) may not be optimal in all scenes; the method can remove degenerate but non-outlier edges when combined with Triangle.
- Trained on a single outdoor scene with thresholded labels derived from COLMAP reconstructions.
Experimental setup
Datasets: PhotoTourism 1DSfM ScanNet
Benchmarks: 1DSfM PhotoTourism ScanNet
Metrics: mean position error position recall@10m position recall@5 position recall@1m number of cameras retained runtime
Baselines: 1DSfM MSAC score CleanNet Triangle filter Oracle filter No filter
Builds on: 1DSfM GLOMAP NeuRoRA DINOv2 TransformerConv
Closest papers by meaning
- Learning Structure-from-Motion with Graph Attention Networks (CVPR 2024 · similarity 0.87)
- Leveraging Camera Triplets for Efficient and Accurate Structure-from-Motion (CVPR 2024 · similarity 0.86)
- Revisiting Global Translation Estimation with Feature Tracks (CVPR 2024 · similarity 0.85)
- Light3R-SfM: Towards Feed-forward Structure-from-Motion (CVPR 2025 · similarity 0.85)
- Cycle-Sync: Robust Global Camera Pose Estimation through Enhanced Cycle-Consistent Synchronization (NeurIPS 2025 · similarity 0.85)
- FAR: Flexible Accurate and Robust 6DoF Relative Camera Pose Estimation (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.