After matching the keypoints calculate the homography matrix. There are lots of tutorials on feature matching and homography estimation, for example here and here.
Now, in the homography matrix, the X and Y scale factor will correspond to the values at [0,0] and [1,1] positions of the matrix.
The translation is at positions [2,0] and [2,1]; skewing factor at [1,0] and [0,1] etc.
Thank you very much. But how about cases with a lot of false matching? Actually, my data is with a lot of repetitions, so I can't identify the correspondences exactly. And this is why I asked this question. Though there is ambiguity, I think the scale factor will not change. So I want to get the factor by feature matching.
The homography estimation will try to get the best pair of SIFT points to match the images. If it doesn't work, you have to find another method to match the images.
Anyway, your problem is not clear; maybe if you add an example image, it will be easier to answer.
See the Computer Vision: Algorithms and Applications book by Richard Szeliski. There is a chapter about 2D transformations:
If you can assume that the two objects follow a similarity transformation, you have:
And you can use findHomography() in a robust scheme to discard the outliers. You should get H that corresponds to this transformation with the last row being [0,0,1].
Or more generaly, the affine transformation with some specific operations:
In this case, you should be able to decompose the affine transformation, e.g. here or here.
Most likely you are dealing with pictures of object taken by different viewpoints. In this case, the induced transformation in the image plane is a perspective transform or homography:
For some specific camera motions, the homography will encode a simplified transformation. But generally this should not be true. Also, the homography holds if:
See this tutorial: Basic concepts of the homography explained with code for additional information.
Why not? Easy approach would be to get the keypoints from the both the image and section them into segments, to prevent false matching. You can start with 2x2 sections. Maybe 3x3 or 3x4 given how big is your image and expected aspect ratio. Now match corresponding section's keypoint based descriptor with each other. Play with the strictness of the matches. Now when you have the matches. Scale is easy to calculate in both direction X and Y, by comparing the distance between the matched keypoints in corresponding images.
Thank you very much. But how can we section the key points? Actually, my data is with a lot of repetitions, so I can't identify the correspondences exactly. And this is why I asked this question. Though there is ambiguity, I think the scale factor will not change.
Asked: 2019-06-03 20:09:12 -0600
Seen: 3,773 times
Last updated: Jun 05 '19
