Question about Affine matrix structure

[Justdjent]

Hi @kevinzakka
First of all, thank you for detailed explanation and implementation. But I have a question about theta matrix.
I want to use spatial transformer network to estimate affine transformation from 1 image to another and then use those Affine matrix with OpenCV function cv2.warpAffine().
I understood that translation is normalized between (-1,1), but how to deal with scale and rotation parameters?

I created a function to convert the matrix to appropriate format, but it don't work:

def tf_to_cv(theta, input_img):
    rows, cols, ch = input_img.shape
    # print(theta.reshape(2,3))
    M = theta.copy().reshape(2,3)
    M[0][0] = 1/M[0][0]
    M[1][1] = 1/M[1][1]
    M[0][1] = 1/M[0][1]
    M[1][0] = 1/M[1][0]
    M[0][1] = M[0][1] * -1
    M[1][0] = M[1][0] * -1
    M[0][2] = -cols * M[0][2]/4
    M[1][2] = -rows * M[1][2]/4
    # print(M[0][2], M[1][2])
    M[0][2] = M[0][2] + ((1 - M[0][0]) * cols/2 - M[0][1]*rows/2)
    M[1][2] = M[1][2] + (-M[1][0] * cols/2 + (1-M[1][1])*rows/2)
    return M

[Pay20Y]

Hey I have the same question with you. How to make the affine matrix computed from OpenCV work? Did you solve it? Thanks! @ @Justdjent

[Pay20Y]

I solved it. Sorry to bother you.

[JordanCheney]

@Pay20Y could you share your solution?

[Pay20Y]

@Pay20Y could you share your solution?

Sorry, it's been a long time, I've forgotten. I use OpenCV function cv2.getAffineTransform to estimate matrix and use STN to transform the image. The code can be found here.

[persuelx]

@Pay20Y Your answer is totally out of line with the requirements.You should see what people ask.