Centering opencv rotation [closed]

edit

I'm having difficulties getting opencv rotations to center.

The rotation must retain all data so no clipping is allowed.

My first test case is using 90 and -90 degrees to simplify the transformation matrix (see https://docs.opencv.org/3.0-beta/doc/...)

I also thought the best way to observe rotations is to use a simple case where the border pixel values are set to observe how the box rotates.

The python code I tried came from Flemin's Blog on rotation (http://john.freml.in/opencv-rotation)

Below is a picture of the original non-rotated image in python. Use the grey point as (4,9) for reference.

Then after running the python script (script below), I get a rotation where it is shifted to the right one column. Note the reference point is at (1,4) when it should be at (0,4)

Below is the Python script. I added width and height offsets to the function to allow me to experiment with offsets to the tx and ty rotation parameters. I found that setting the width offset to 1 made the 90 degree rotation case match Matlab, but it didn't help -90.

UPDATE 1/19 9AM: I tried setting offset = -0.5 in the function rotate_about_center() below and the 90 and -90 degree rotations center as expected. For a 10x10 image, the reasoning why this may work is that the center point defined by (cols/2, rows/2) is not (5,5), but rather (4.5, 4.5). The same logic applied to a 11x11 image: the center is not (5.5,5.5) but rather (5,5). Rotations at 45 and -45 don't center - meaning they visually don't look centered in the box computed of size nw x nh. So I think I understand why a "center" equal to (cols/2 - 0.5, rows/2 - 05) works but a center of (cols/2, rows/2) does not, however, most examples I've found do not subtract the 0.5.

import cv2
from matplotlib import pyplot as plt
import functools
import math

bwimshow = functools.partial(plt.imshow, vmin=0, vmax=255,
                             cmap=plt.get_cmap('gray'))

def rotate_about_center(src, angle, widthOffset=0., heightOffset=0, scale=1.):
    w = src.shape[1]
    h = src.shape[0]

    # Add offset to correct for center of images.
    wOffset = -0.5 
    hOffset = -0.5

    rangle = np.deg2rad(angle)  # angle in radians
    # now calculate new image width and height
    nw = (abs(np.sin(rangle)*h) + abs(np.cos(rangle)*w))*scale
    nh = (abs(np.cos(rangle)*h) + abs(np.sin(rangle)*w))*scale
    print("nw = ", nw, "nh = ", nh)
    # ask OpenCV for the rotation matrix
    rot_mat = cv2.getRotationMatrix2D((nw*0.5 + wOffset, nh*0.5 + hOffset), angle, scale)
    # calculate the move from the old center to the new center combined
    # with the rotation
    rot_move = np.dot(rot_mat, np.array([(nw-w)*0.5 + widthOffset, (nh-h)*0.5 + heightOffset,0]))
    # the move only affects the translation, so update the translation
    # part of the transform
    rot_mat[0 ...