Questions about _get_transform_matrix

[zhangtingyu11]

I have some questions about function in this file. I have learned about cv2.getAffineTransform which needs 3 corresponding point pairs. The first pair in this function is the center point of two pictures. But I don't know why the second points pair should be that in the code. I think the src_dir and dst_dir should be [0, src_h * -0.5] and [0, dst_h * -0.5]. Can anyone explain why the height dimension should minus the 0.5*width?

[zhangtingyu11]

I have changed src_dir and dst_dir to what I think, and retrain the SMOKE model, and the training process does not converge. But I still don't understand why should minus the 0.5*width. Can anyone help me?