Image registration

Question 1

Describe ICP process

Answer 1

First stage: identifying closet points in a model to each data point that has been observed
Second stage: finding the best (least squ sense) rigid body transformation relating these point sets
Set of data is then transformed using the transformation matrix, the process of matching closet points is repeated until the change in mean squ error is below a predetermined threshold

Question 2

Describe RANSAC process aka geometric verification

-> image retrieval

Answer 2

1) point pairs are sampled from image A and B and are assumed to define correspondence
through finding patches around distinctive points and then paired with their best matching points (eg: corners)
2) rigid body transformation is estimated through the correspondence
3) other point pairs are then transformed in A not using the estimated transformation
4) compare the transformed point in A with points in B
5) find the focus points/inliers
6) iterate through different combinations of A and B
7) take the transformation that results in the greatest number of inliers

Question 3

Problem of RANSAC

Answer 3

• requires a variable number of iterations to get the best solutions
• computationally intensive
• if unpaired feature points are numerous, higher chances of being incorrect

Question 4

Describe Voxel similarity process

Answer 4

-for matching two images or volumes with no known homologous points
-image pixel or voxel values themselves are used to estimate geometric transformation
-having some measure of similarity between two images
then iterate to find the geometric mapping which takes a point in image A to image B
-then do it to all of the common image space for A and B