Point registration - How to get the eigenvectors to have a consistent direction?
12 ビュー (過去 30 日間)
I'm trying to replicate the algorithm of the attached paper. It's about feature-based correspondence based on the eigenvectors. I'll summarize:
We have a point cloud/shape (as in Figure 2, which I'm trying to replicate) and create a matrix H (adjacency of the points) which describes the relation of the intradistances (not interdistances) in an image. From this matrix we calculate the eigenvectors and values. They have to be reordered from big to small and the sign of the vector adapted, so that they have a consistent direction. The rows describe the features. Finally we calculate square of the euclidean distance between the features to get the matrix Z. The smallest number in every row tells us, which point corresponds to which.
Now I get the same numbers for the eigenvectors, but not always the right sign. For the matrix U1 of the eigenvectors of the first body, the second column should be multiplied by -1. For the matrix U2 of the eigenvectors of the second body, the second and fourth column should be multiplied by -1.
I tried to find out if this is somehow apparent when calculating the angles between eigenvectors (see the code) but the result (see the matrix bellow the code) doesn't tell me anything. Maybe you have a better eye.
for i = 1:minNofP;
for j = 1:minNofP;
ThetaInDegrees.CosTheta(i,j) = max(min(dot(U1(:,i),U2(:,j))/(norm(U1(:,i))*norm(U2(:,j))),1),-1)
ThetaInDegrees.Angle(i,j) = real(acosd(ThetaInDegrees.CosTheta(i,j)));
The problem is, that without the adaptation of the sign, the algorithm doesn't work. This may be a "just" a mathematics problem, but I am not able to solve it and I know that here there are a lot of people a lot more knowlegdeable (and smarter) than me.
There are many things attached:
1) The pdf of the article, in case you want to thow an eye on it. The consistent direction thing is explained on the third page on the right side before the ecuation.
2) Adjacency_V2 is an adapted version of the adjacency function from https://www.mathworks.com/matlabcentral/fileexchange/29227-eigen-function-of-the-laplacian. I adapted it to use my euclidean distance function. It should not be a problem, it has worked fine for me.
3) TruncateToMinNumber is a function to truncate a matrix to the smaller number of eigenvectors. It doesn't have an effect on this example, as the matrices have the same amount of eigenvectors and so they stay the same.
4) Correct EV have the correct eigenvectors given on the paper. The algorithm works with it. I've used it to compare.
5) DistanceBetweenF is my function for the euclidean distance between the features.
6) Features_EV_EW_v2 is my code. I've written %IGNORE on the lines you don't need but I dind't want to erase.
I tried to make the question as complete as possible but so summarized that someone actually reads it. I'm very sorry if I forgot some function or something relevant for you to be able to help me. I will edit it in as soon as I know about it.
As always thank you very much.
回答 (1 件)
Dana 2020 年 9 月 25 日
This isn't my area of expertise, but one simple "orientation" test is as follows. Consider two scalars a and b. One way to check whether these scalars are of the same sign (i.e., "oriented in the same direction") is to compare with . If the former is larger they are of the same sign, while if the latter is larger they are of opposite signs. We can extend this principle to the case of vectors a and b by saying that a and b are oriented in the same direction if and only if . Otherwise, they are oriented in opposite diretions (in which case we would want to flip one of them).