Structure from Motion

Traditional photogrammetry required carefully arranged photos and manually measured control to compute 3D coordinates, demanding expertise and time. SfM automates this: it finds matching features across many overlapping, freely taken photographs and solves for both the camera positions and the scene geometry simultaneously by bundle adjustment, so a non-specialist with a camera can produce a model. Compared with laser scanning, which directly measures distances and excels on textureless or complex surfaces, SfM is far cheaper and yields rich photographic texture but depends on the subject having visible surface detail and on good capture geometry for accuracy. The two are complementary, and many projects combine them.

SfM recovers depth the way binocular vision does — by seeing the same point from different viewpoints. Each physical point must appear in several photographs taken from sufficiently different positions for the geometry to be solved reliably; if overlap is too low or angles too similar, features cannot be matched across enough images and the reconstruction develops holes, weak geometry, or systematic distortion such as doming. Recommended practice is therefore heavy overlap (often sixty to eighty percent or more) and convergent, all-around coverage, with sharp, well-lit images. The discipline of capture is the single biggest determinant of model quality, which is why field protocols emphasize photo planning before processing.

Bundle adjustment reconstructs the scene only up to an arbitrary scale and orientation, so an extra step anchors the model to reality. Scale bars of precisely known length placed in the scene fix the model's size, and ground control points whose real-world coordinates are measured with a total station or GNSS receiver let the software compute a similarity transformation that rotates, scales, and translates the model onto the survey datum. The fit is checked by examining residuals at the control points, which quantify accuracy. Once georeferenced, as Conolly and Lake describe for spatial data generally, the model's orthophotos and elevation models can be measured and analyzed in GIS alongside other site information.