3.3. Precision and applicability
In general terms, the accuracy of a photo-based model depends
upon the scale and resolution of the input images, the distribution
and accuracy of control data (whether ground control points, scale
measurements or camera positions), the precision and distribution
of matched image points, and the network geometry, which includes
the number of photos, how much they overlap and how
convergent the views are. In close-range photogrammetry (e.g., as
often used for high accuracy engineering applications), where image
networks are usually multi-image and highly convergent, the
strength of a network can be described by its relative network
precision, which is a ratio of the mean 3D point uncertainty estimate
to the longest dimension of the network. For a given image
measurement precision, a stereo image pair with only two nearparallel
images would represent a weaker network than a convergent
multi-image arrangement. For projects using digital SLR imagery
covering sub-meter to kilometer scales processed with an
SfM-based approach, James and Robson (2012) estimated relative
precision ratios of ~1:1000 or greater. These ratios were shown to
be similar to those of theoretical estimates for stereo photogrammetry,
but approximately an order of magnitude poorer than
equivalent theoretical estimates for close-range (convergent)
networks.