Triangulation-sensitivity-guided view selection for structure from motion
Year: 2026
Type: article
Abstract: Structure from Motion (SfM) is a core technique for three-dimensional (3D) reconstruction and supports artificial intelligence (AI)-enabled engineering applications, including industrial measurement and robotic navigation. Traditional view-selection strategies mainly rely on image similarity or feature-matching strength, while the influence of view geometry on triangulation stability is often handled using fixed heuristics. This paper contributes an interpretable, geometry-driven framework for automated view selection in SfM. The framework implements AI-based decision making through a directed triangulation-sensitivity matrix constructed from preliminary camera poses and sparse 3D points by measuring the 3D displacement induced by a standardized one-pixel perturbation. Adaptive candidate-view ranking, bidirectional verification, and view-graph completion are then integrated into the incremental reconstruction process to select reliable image relationships and balance reconstruction accuracy and efficiency. Experiments on the Technical University of Munich (TUM), Technical University of Denmark (DTU), and Alcatraz West Side Gardens datasets cover indoor red-green-blue and depth (RGB-D) sequences, object-centered multi-view scenes, and large-scale outdoor image collections. Across the evaluated scenarios, the method attains the lowest mean reprojection error in 11 of 14 TUM configurations, produces lower mean reprojection error than Exhaustive matching on all 20 DTU scans, and reconstructs 51,831 sparse 3D points on West Side. Ablation, parameter-sensitivity, pixel-perturbation, scalability, and runtime analyses further characterize the effectiveness, robustness, and computational trade-offs of the proposed method.
Institutions: Jiangnan University, Brunel University of London
Language: English
Topic: Advanced Vision and Imaging
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,200
OpenAlex ID: W7214537491