PDF version of this document

Method

Given a set of non-rigidly registered images, we can build a generative appearance model that captures both the shape and intensity variation across the set, by performing a joint statistical analysis of the warp fields x resulting from the NRR and the shape-free intensity patterns g measured in the frame of the mean shape.

3#3 = 4#4

We can use the generative property of the model to synthesise a large set of images, 5#5 which, if the model is good, should form a cloud that overlaps the cloud of training images.

2#2

We define the Specificity 6#6 of the model as 7#7 where 8#8 is the size of the training set, 9#9 is a measure of distance between images, 10#10 is the 11#11 training image, and 12#12 is the minimum over 13#13 (the set of synthetic images). 6#6 will be small if the two clouds fully overlap.

3#3 = 14#14