3.2.4. Densifying the measurements: multi-view stereo
With known camera models and orientations, a multi-view
stereo algorithm will produce a dense point cloud representation
of the surface. Typically, this technique will be implemented as a
systematic search over a pixel grid to identify best matches between
images, with the results providing significantly more 3D
points, having greater precision than the feature matching of the
initial SfM step. Multi-view stereo is particularly intensive
computationally if the full image collection is processed simultaneously.
However, most photo-based 3D reconstruction programs
and algorithms have the option to subset image collections (e.g.,
Furukawa et al., 2010) or to adjust the grid-cell size at which multiview
stereo is performed so as to manage the resolution and time
required to produce the resultant dense point cloud.