仿射变换矩阵

Question

I'm trying to perform image stabilization on a set of images. I have identified matching pairs of corner points in each image and am trying to calculate the affine transformation matrix for each set of matching pairs (source vs. destination). When I do this, though, using a least squares method, the resulting transformation matrix produces images that are not stabilized. I am following this process:

source_points = [[219, 187], [221, 387], [347, 31], [135, 311], [296, 615], [86, 417]]
dst_points = [[221, 187], [222, 387], [348, 30], [137, 310], [299, 615], [88, 417]]

transform_mat, res, rank, s = np.linalg.lstsq(source_points, dst_points)

#looping through all dst_coordinates in a blank array of size = source image
transformed_coor = np.dot(transform_mat,dst_coordinate)

After looping through the coordinates in the blank array, I grab the interpolated value of the transformed_coor from the source image and place that value at the dst_coordinate in the blank array. When the process completes I get a transformed image but it is not stabilized when I compare it with the source image.

Is this process correct? Am I missing a step?

Answer 1

I think your point correspondences source_points and dst_points are not matching well. For this very input I computed the affine transformation matrix

T = [0.9997   -0.0026   -0.9193
     0.0002    0.9985    0.7816
          0         0    1.0000]

which leads to individual transformation errors (Euclidean distance) of

errors = [0.7592    1.0220    0.2189    0.6964    0.4003    0.1763]

for the 6 point correspondences. Those are relatively large, especially when considering the distances between source_points and dst_points of

move_dist = [2.0000    1.0000    1.4142    2.2361    3.0000    2.0000]