how to stitch two images with overlapped area?

hi

i am using SIFT algorithm with vl_sift (and other functions) to find matched points in two images, that has overlapping area. now that i have the match points, how can i transform the second image and stitch it to the first one. from what i read in different papers, i conclude that i have to create mask to do this, but then what to do with the mask?

here is my code: (of course incomplete, i need now to stitch these two pics together)

f1 = imread('IMG_0148.jpg');
f2 = imread('IMG_0149.jpg');
f1 = im2single(rgb2gray(f1));
f2 = im2single(rgb2gray(f2));
[h1,d1] = vl_sift(f1) ;
[h2,d2] = vl_sift(f2) ;
[matches, scores] = vl_ubcmatch(d1, d2) ;

Thanks!

Answer 1

[Image Analyst]

See Chapters 9 and 10 of Richard Szileski's book: http://szeliski.org/Book/ which discuss this topic.

[Mahnaz Pariyan]

thanks. it's great!

Answer 2

[Omair]

I'm assuming you are trying to create a panorama from these two images. For that you will need to calculate the homography and transform one image into the plane of the other image before you can stitch them together.

[Mahnaz Pariyan]

exactly that's what i want to do, but my problem is how to transform it into the plane of the other image.

[Omair]

Have you calculated the homography matrix yet? I think that would be needed to transform one image coordinates into the plane of the other.

[Mahnaz Pariyan]

i just did that, now what's next with it?

[Omair]

Now you need to transform your pixel coordinates to the pixel coordinates of the other image plane by multiplying with the homography matrix, not the pixel values but the pixel coordinates. This will tell you where the particular pixel value belongs in the other image plane.

[Mahnaz Pariyan]

here i got some problems:

in calculating the homography matrix ,in the code i used, it is written that the input points must be 2xN matrices, if they are same as i have as "matches" and "scores", then one of them is 2x202 and the other one is 1x202. I thought by zero-padding the problem would be solved, and i added another row to the "scores" matrix.

now for applying the homography matrix on the image coordinates, i get no error actually, but the result is a 3x7596 matrix, which seems to be wrong.

Answer 3

[Omair]

Have a look at these slides ( http://www.gris.tu-darmstadt.de/teaching/courses/ss13/cv1/slides/l10-geometry.pdf), slide 15 onwards . For calculating the homography matrix you have 8 unknows. You need 4 feature point correspondences between your two images to solve for those 8 unknows. From those you calculate the homography matrix.

[Mahnaz Pariyan]

thanks, i'm reading it.