kNN and Entropy in Registration and Model Assessment
Date: Thursday, December 15 @ 08:16:31 GMT
Topic: Implementation

REVIOSULY we used model indices, which we called Generalisation and Specificity, to assess the quality of appearance models, as well as the quality of non-rigid registration. We have now identified a valuable surrogate to these indices: Shannon's entropy. Some work by Hero et al. is encouraging the use of entropic measures to assess (dis-)similarity of graphs. This is practically used as non-rigid registration similarity measures -- somewhat reminiscent of mutual information (MI).

We intend to see if an entropic measure of clouds overlap suprasses the performance of Generalisation and Specificity. We also consider image distances that are based on K nearest neighbours (kNN) or the nearest match to a pixel intensity, a map of which is shown below. Since it takes around 20 minutes to generate each of the images below, we consider this to be highly impractical. To run just a single such model evaluation, we would need over 60,000 hours of computer power. And this is 2-D only...

Extension of our approach to 3-D is foreseen nonetheless. It will probably use the methods which require only a couple of hours of computation in 2-D. Resolution 'pyramiding' (coarse-to-fine approach) can assist in terms of speed.

Top: original image; Bottom: nearest match to pixel of greyscale value 20, 60, 100 (left-to-right) for each of the other pixels in the image

This article comes from MARS - Models of Appearance, Registration and Segmentation

The URL for this story is: