Distribution of residuals from each orthophoto (fig. 9) reveals
similar shift pattern among LPS and Photoscan, while the
proposed method exhibits smaller and randomly distributed
residuals. This may be due the nature of the differential
rectification process which inevitably stretches the photograph
on occluded areas. At the same time, the photogrammetric
process may perform worst as the layout of the photos isn’t a
proper one, which significantly reduces overall accuracy.
Computer-vision based methods with much more tie points
perform better in this aspect. The proposed method seems more
accurate despite the fact that the registration process was based
on points from the point cloud instead on the photos
themselves, hence less precise