Structure from Motion

Structure from Motion (SfM) recovers camera positions and a 3D scene from overlapping photographs or video, enabling mapping, measurement, and novel-view rendering without a prebuilt model. Classical pipelines detect and match features, assemble image view graphs, triangulate points, and refine poses and geometry with bundle adjustment; incremental and global methods handle different scales and capture patterns. Newer approaches use dense multi-view matching, monocular depth, or 3D foundation-model predictions, often combining these learned priors with geometric optimization. Methods are tested on settings such as ETH3D and IMC Phototourism, but difficult scenes with repeated structures, motion, low light, or weak overlap expose a continuing tension between robustness, geometric accuracy, and computational cost.