Research map / 3D vision / Depth and stereo
Learning Global Camera Poses from Noisy View-Graphs for Structure from Motion
Fadi Khatib, Meirav Galun, Ronen Basri
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use 3 GNN layers instead of 2 (ablation) | VGPA · MegaDepth validation, before final BA | pose AUC@30: 0.346 -> 0.389 | Improved |
| use edge-conditioned message passing instead of GATv2 attention (ablation) | VGPA · MegaDepth validation, before final BA | pose AUC@30: 0.364 -> 0.389 | Improved |
| add test-time fine-tuning to pretrained model (ablation) | VGPA · MegaDepth validation, before final BA | pose AUC@30: 0.358 -> 0.389 | Improved |
| add pretraining before fine-tuning (ablation) | VGPA · MegaDepth validation, before final BA | pose AUC@30: 0.352 -> 0.389 | Improved |
| train the network rather than random init (ablation) | VGPA · MegaDepth validation, before final BA | pose AUC@30: 0.012 -> 0.358 | Improved |
| add view re-integration post-processing (ablation) | VGPA · MegaDepth (mean) | registered images: 327 -> 348 (rot 0.95 -> 1.36) | Mixed |
| replace exhaustive graph with sparse MegaLoc top-30 retrieval graph (ablation) | VGPA · 1DSfM | rotation error: Piazza del Popolo 5.37 -> 0.57; Yorkminster 0.43 -> 6.32 | Mixed |
| estimate intrinsics jointly in BA instead of known intrinsics (ablation) | VGPA · MegaDepth scene 0048 | rotation error: 0.10 -> 0.38 | Hurt |
| estimate intrinsics jointly in BA instead of known intrinsics (ablation) | VGPA · BlendedMVS scene0 | rotation error: 0.149 -> 0.122 | No clear effect |
| use VGPA instead of RESfM deep factorization | pose estimation · MegaDepth (mean) | AUC@30: 0.910 -> 0.927 | Improved |
| use VGPA instead of GLOMAP | pose estimation · 1DSfM runtime (mean) | total time (min): 18.2 -> 2.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- RESfM — RESfM: Robust Deep Equivariant Structure from Motion (ICLR 2025) · better on 4 of 4 shared results
- VGGT — VGGT: Visual Geometry Grounded Transformer (CVPR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Learning Scene Coordinate Reconstruction from Unposed Images via Pose Graph Optimization (CVPR 2026 · similarity 0.91)
- Learning to Filter Outlier Edges in Global SfM (CVPR 2025 · similarity 0.89)
- Light3R-SfM: Towards Feed-forward Structure-from-Motion (CVPR 2025 · similarity 0.87)
- Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.87)
- Learning Structure-from-Motion with Graph Attention Networks (CVPR 2024 · similarity 0.87)
- LEAP: Liberate Sparse-View 3D Modeling from Camera Poses (ICLR 2024 · similarity 0.87)
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.
