PDF version of this entire document
One obvious way to test the performance of a registration algorithm is to apply it to some artificial data where the correct correspondence is known. Such test data is typically constructed by applying sets of known deformations (either spatial or textural) to real images. This artificially-deformed data is then registered, and evaluation is based on comparing the deformation fields recovered by the registration algorithm with those that were originally applied [6,7]. This approach can be used to compare the performance of different NRR algorithms, but since it relies on the creation of artificial test data, cannot be applied in-line. Also, the validity of the approach depends on the ability to construct artificial deformations which mimic the variability found in real images of a given type, which is difficult to guarantee.