photography is typically collected with 50e60% overlap between
adjacent images, under near-parallel viewing conditions (e.g.,
Krauss, 1993; Abdullah et al., 2013). SfM-based 3D reconstruction
methods also require overlapping images but, because they can
operate on unordered collections of photographs, the overlap
requirement is best considered in terms of coverage and angular
change between overlapping images. In terms of coverage, every
surface that will be reconstructed needs to be covered by at least 2
images taken from different positions, and preferably more (Fig. 3).
Increasing angles of convergence between overlapping images will
tend to increase reconstruction accuracy up to a point, but will
eventually prevent matching due to the surface texture appearing
too dissimilar in images from different directions. Moreels and
Perona (2007) found that popular feature detectors used for automated
image matching did not perform well with angular changes
greater than 25e30 between images. Thus, while angular changes
between photos can increase the accuracy of reconstructed 3D
surfaces, differences should be limited to 10e20 for overlapping
photos.