Source-linked AI summary
Robust Global Structure-from-Motion via View Graph Pruning
Jiamin Xu, Lixing Yao, Weichen Dai, Renshu Gu, Zunjie Zhu, Weiwei Xu, Gang Xu
TL;DR
Erroneous view-graph edges from visually ambiguous matches can undermine global SfM reconstruction. The paper prunes such edges using internally consistent subgraphs and RANSAC, achieving improved robustness and novel-view synthesis quality.
Problem
Global SfM is highly sensitive to spurious view-graph edges from false feature matches, especially in scenes with repetitive structures.
Method
The method recursively partitions the view graph into pose-consistent subgraphs, estimates local poses, and uses RANSAC with relative pose and scale consistency to prune inter-subgraph edges before global optimization.
Results
The method achieves competitive or superior camera-pose accuracy across sequential and unordered datasets, with particularly strong improvements in visually ambiguous scenes and improved 3DGS novel-view synthesis.
Takeaways & Limitations
Early removal of unreliable connections prevents error propagation during rotation and translation averaging, producing more accurate and geometrically consistent reconstructions.
Takeaways & Limitations
Performance may remain limited when repetitive regions are too large or densely distributed, and the approach depends on hyperparameters that may not be optimal.
Abstract
from arXiv · showhide
Structure-from-Motion (SfM) aims to estimate camera poses and reconstruct 3D structures from a collection of unordered images. Compared with incremental SfM, global SfM achieves better scalability by jointly estimating camera poses based on a view graph constructed from pairwise correspondences. However, its performance is highly sensitive to erroneous edges caused by visually ambiguous matches, which may lead to incorrect camera registration and reconstruction artifacts. In this work, we propose a subgraph-guided view graph pruning framework for robust global SfM. Our key idea is to exploit the internal consistency of reliable subgraphs to identify and remove unreliable connections. Specifically, we first partition the view graph into locally consistent subgraphs and perform global SfM within each subgraph to obtain reliable camera poses. We then apply RANSAC-based edge pruning across subgraphs to remove inconsistent edges, and finally perform global SfM on the refined view graph. Extensive experiments on ambiguous, sequential, and unordered image datasets demonstrate that our method improves the robustness of global SfM under challenging conditions. Further evaluation with neural rendering shows that the improved camera estimation leads to higher-quality novel view synthesis results.
1 Introduction
SfM estimates camera poses and sparse 3D structure, while global methods improve scalability but remain vulnerable to erroneous view-graph edges. The paper addresses this vulnerability with recursive subgraph partitioning and inter-subgraph pruning before final global reconstruction.
- SfM simultaneously estimates camera poses and reconstructs sparse 3D structures from multiple images.
- Global SfM estimates camera parameters jointly from a view graph, improving scalability relative to incremental approaches.
- Erroneous edges from false matches, especially in repetitive or visually ambiguous regions, can cause incorrect global registration and reconstruction artifacts.
- The method recursively partitions the view graph into internally consistent subgraphs and performs global SfM within each subgraph.
- RANSAC-based pruning removes inconsistent inter-subgraph edges before final global SfM, preventing unreliable connections from propagating errors during averaging.
- The method achieves competitive or superior camera-pose accuracy, particularly over GLOMAP in visually ambiguous scenes, while retaining comparable global-SfM efficiency.
2 Related Work
Prior SfM research balances incremental robustness against global scalability, while addressing outliers through geometric verification, robust optimization, and graph refinement. This work instead prunes inter-subgraph edges and directly produces the final reconstruction through global SfM.
- Incremental SfM progressively adds cameras and 3D points with repeated optimization, typically achieving high accuracy but incurring computational expense and possible drift.
- Global SfM jointly estimates camera poses through rotation and translation averaging, improving efficiency and scalability but remaining sensitive to incorrect correspondences.
- Traditional outlier rejection uses RANSAC with fundamental or essential matrix constraints, while other methods reject points or images using reprojection errors.
- View-graph partitioning has been used to reconstruct components independently and align or merge partial models hierarchically.
- This method refines the view graph by applying RANSAC-based edge pruning between subgraphs and obtains the reconstruction directly through global SfM rather than merging partial reconstructions.
3 Proposed Method
The method reconstructs a view graph, recursively partitions it into pose-consistent subgraphs, prunes inconsistent inter-subgraph edges, and reruns global estimation on the refined graph.
- View Graph Reconstruction: GLOMAP constructs a view graph from verified feature correspondences, with images as nodes and edges representing image pairs containing sufficient matching inliers.Two-view geometry estimates relative poses from the surviving inlier correspondences.
- Global Camera Pose Estimation: Global estimation jointly recovers absolute camera poses through rotation averaging, global positioning, and global bundle adjustment.Rotation averaging estimates absolute rotations from relative rotations, while global positioning jointly estimates camera positions and 3D points.
- Subgraph Partition: The graph is recursively partitioned with Louvain community detection into densely connected subgraphs until intra-subgraph camera pose consistency is achieved.Partitioning uses inlier-correspondence counts as edge weights and evaluates loop consistency plus relative-to-absolute pose agreement.
- Inter-Subgraph Edge Pruning: For inter-subgraph edges, PnP estimates poses across local coordinate systems, and RANSAC aligns subgraph scales and rejects inconsistent connections.The scale factor is estimated from edges connecting two subgraphs, using translation discrepancies among sampled relative poses.
4 Experiments
Experiments evaluate the method on ambiguous, sequential, and captured datasets using camera-pose accuracy, qualitative geometry, ablations, and runtime. The method performs strongly under ambiguous matching while retaining global-SfM efficiency.
- Ambiguous image datasets: The method robustly recovers all four ambiguous benchmark scenes, whereas competing methods show incorrect registrations, incomplete structures, or violated epipolar constraints.Qualitative evaluation verifies reconstructed point clouds, camera trajectories, and epipolar-line consistency.
- Ambiguous image datasets: 90.1, 93.7, and 96.4 are the highest average AUC values at 3°, 5°, and 10° across the four ambiguous datasets.On Cup, Doppelgangers++ remains best, while the proposed method ranks second with 66.4, 78.4, and 87.6.
- Novel view synthesis: The estimated camera poses consistently improve novel view synthesis over traditional SfM approaches when evaluated with 3DGS.The evaluation uses PSNR on held-out testing images.
- Sequential image datasets: On Family and Auditorium, the method substantially outperforms COLMAP, although Doppelgangers achieves the best performance on both sequences.The comparison includes GLOMAP, COLMAP, Doppelgangers, Doppelgangers++, and GraphSfM.
- Sequential image datasets: The proposed method achieves the best AUC scores on Exhibition Hall and Office, and ranks second to Doppelgangers on Old Computer.These results come from an additional benchmark of three sequential sequences.
- Ablation study: Recursive subgraph partitioning is critical: removing it reduces AUC@10° from 84.8 to 14.6.A single Louvain partition is insufficient to isolate geometrically inconsistent regions reliably.
- Computational time: The method runs over three times faster than COLMAP and approximately five times faster than Doppelgangers on Courthouse, while slightly slower than vanilla GLOMAP.Its global-SfM design preserves substantial runtime advantages on longer sequences.
5 Conclusion
The paper presents a robust global SfM method for large scenes with repetitive or duplicate structures, while identifying remaining limitations and future directions.
- 5 Conclusion: The framework combines subgraph partitioning with intra-subgraph pose consistency to support robust global SfM.It also uses RANSAC-based filtering to remove inconsistent edges across subgraphs.
- 5 Conclusion: The method is designed to handle large scenes containing repetitive or duplicate structures.
- 5 Conclusion: Errors may remain when repetitive regions are too large or densely distributed, and several hyperparameters may be suboptimal.Future work targets learning-based techniques and advanced matching strategies.