PDF version of this document

Method

We automatically build an appearance model from the correspondences identified by a registration. We then synthesise images from the model and perform a cross-imageset comparison to assess model/registration quality.

3#3 = 4#4

• Generate synthetic images from the AM
• Compare training images with synthetic images
  • images can be vectorised
  • embed images in hyperspace
  • estimate clouds overlap

2#2

3#3 = 5#5

Given 6#6 training images, we use the generative property of the model to synthesise a large set of images, 7#7. We can then define Specificity 8#8:

9#9 (1)

where 10#10 is a measure of distance between images, 11#11 is the 12#12 training image, and 13#13 is the minimum over 14#14 (the set of synthetic images).