Structure from Motion (SfM) is the computer-vision approach to photogrammetry that made 3D reconstruction ubiquitous. Given many overlapping photographs — from a drone, a handheld camera, even a phone — SfM automatically finds matching features, solves camera positions and geometry simultaneously, and then densifies the result into point clouds, meshes and orthomosaics. It is the engine inside virtually every drone-mapping package.
What the algorithm does
SfM detects thousands of distinctive features per image, matches them across views, and solves a joint optimisation — bundle adjustment — recovering both the 3D structure of the scene and the motion (positions, orientations, calibration) of the camera, with no prior survey of either.
Multi-view stereo then correlates pixels between adjusted images to build dense clouds rivalling LiDAR density on well-textured surfaces, from which meshes, DSMs and orthos follow.
From reconstruction to measurement
Raw SfM is internally consistent but floats in arbitrary scale and position until georeferenced. Ground control points — or RTK/PPK camera positions — anchor the block to real coordinates, and independent checkpoints prove the result. With disciplined control, drone SfM routinely achieves centimetre-level accuracy.
Without that discipline, outputs look superb and measure poorly: the doming and scale errors of uncontrolled blocks are the classic trap of push-button mapping.
Strengths, limits and complements
SfM excels on textured, static, well-lit scenes: quarries, construction sites, facades, stockpiles. It struggles on water, glass, uniform surfaces and moving vegetation, and it cannot see ground beneath canopy — the boundary where UAV LiDAR takes over.
In practice the technologies pair: SfM for photorealistic surfaces and orthos, LiDAR for penetration and geometry, unified in one site model.