rgov's profile - activity

I have a stereo camera system with two different cameras, with different focal lengths, optical center points, and image resolutions. They are positioned horizontally and the relative rotation is negligible.

I've been given the intrinsic matrices for each camera, their distortion coefficients, as well as the rotation matrix and translation vector describing their relationship.

I want to rectify a pair of photos taken by the cameras at the same time. However, the results have been complete garbage.

I've first tried ignoring that the image resolutions are different, and using cv2.stereoRectify then cv2.initUndistortRectifyMap then cv2.remap. Since this didn't work, I added a preprocessing step to scale both images to the same dimension. The algorithm is now:

1. Remove distortion from each image using cv2.undistort
2. Scale the images to the same width and height with cv2.resize
3. Transform the focal length and optical center points of the camera matrices accordingly (per this answer)
4. Perform cv2.stereoRectify with the new camera matrices and zero distortion
5. Compute the rectification map with cv2.initUndistortRectifyMap for each camera
6. Apply the rectification map with cv2.remap on each image

However again the output is garbage. I've re-read the code to make sure I didn't make any copy-paste errors, have compared with similar implementations, and consulted the relevant chapters in the "Learning OpenCV 3" book. I've written out image files at each step to make sure the undistortion and scaling are correct.

Are there any sanity checks I can do to make sure that the camera matrices I'm receiving are correct?

# Undistort the images without rectifying them
left_ud = cv2.undistort(left, left_camera_matrix, left_distortion_coeffs)
right_ud  = cv2.undistort(right, right_camera_matrix, right_distortion_coeffs)

# Now scale the images to the same width and height
mw = max(left_ud.shape[1], right_ud.shape[1])
mh = max(left_ud.shape[0], right_ud.shape[0])

left_s = cv2.resize(left_ud, (mw, mh), interpolation=cv2.INTER_CUBIC)
right_s  = cv2.resize(right_ud, (mw, mh), interpolation=cv2.INTER_CUBIC)

# Adjust the matrices... oops make sure I do (transform)*(camera) not (camera)*(transform)
left_camera_matrix = np.array([
  [float(mw) / left_ud.shape[1], 0, 0],
  [0, float(mh) / left_ud.shape[0], 0],
  [0, 0, 1]
]).dot(left_camera_matrix)
right_camera_matrix = np.array([
  [float(mw) / right_ud.shape[1], 0, 0],
  [0, float(mh) / right_ud.shape[0], 0],
  [0, 0, 1]
]).dot(right_camera_matrix)

# Clear the distortion coefficients
left_distortion_coeffs = right_distortion_coefficients = np.zeros((1,5))


# Rectify both cameras
R1, R2 = np.zeros((3, 3)), np.zeros((3, 3))
P1, P2 = np.zeros((3, 4)), np.zeros((3, 4))

_, _, _, _, _, left_roi, right_roi = cv2.stereoRectify(
  left_camera_matrix, left_distortion_coeffs,
  right_camera_matrix, right_distortion_coeffs,
  (mw, mh),
  R, T,
  R1, R2, P1, P2,
  flags=cv2.CALIB_ZERO_DISPARITY,
  alpha=0.0  # tried different values here
)

# Now compute the rectification maps and apply them
map1, map2 = cv2.initUndistortRectifyMap(
  left_camera_matrix, left_distortion_coeffs,
  R1, P1[:,:3],
  (mw, mh),
  m1type = cv2.CV_32FC1
)

left_out = cv2.remap(left_s, map1, map2, cv2.INTER_LINEAR)
cv2.rectangle(left_out, (left_roi[0], left_roi[1]), (left_roi[0]+left_roi[2], left_roi[1]+left_roi[3]), (0, 255, 0), thickness=3)

map1, map2 = cv2.initUndistortRectifyMap ...

I have two images, src and dst. I'm trying to perform some transformations on src to make it align better with dst.

One of the first transformations I'm applying is a perspective transform. I have some landmark points on both images, and I'm assuming that the landmarks fall on a plane and that all that has changed is the camera's perspective. I'm using cv2.findHomography to find the transformation matrix which represents the change in the camera.

However, if I then apply this transformation to src, some of the image might be transformed outside of my viewport, causing the image to be cropped. For instance, the top left corner (0, 0) might be transformed to (-10, 10), which means this part of the image is lost.

So I'm trying to perform the transformation and get an uncropped image.

I've played around with using cv2.perspectiveTransform on a list of points representing the corners of src, and then getting the cv2.boundingRect which tells me the size of the array I need to store the uncropped image. But I can't figure out how to translate the image so that none of its points get transformed out of bounds.

If I translate src before I apply the transformation, then I think I've "invalidated" the transformation. So I think I have to modify the transformation matrix somehow to apply the translation at the same time.

An answer on StackOverflow from Matt Freeman, on a question titled "OpenCV warpperspective" (I cannot link to it due to stupid karma rules), seemed promising, but didn't quite work.

# Compute the perspective transform matrix for the points
ph, _ = cv2.findHomography(pts_src, pts_dst)

# Find the corners after the transform has been applied
height, width = src.shape[:2]
corners = np.array([
  [0, 0],
  [0, height - 1],
  [width - 1, height - 1],
  [width - 1, 0]
])
corners = cv2.perspectiveTransform(np.float32([corners]), ph)[0]

# Find the bounding rectangle
bx, by, bwidth, bheight = cv2.boundingRect(corners)

# Compute the translation homography that will move (bx, by) to (0, 0)
th = np.array([
  [ 1, 0, -bx ],
  [ 0, 1, -by ],
  [ 0, 0,   1 ]
])

# Combine the homographies
pth = ph.dot(th)

# Apply the transformation to the image
warped = cv2.warpPerspective(src, pth, (bwidth, bheight), flags=cv2.INTER_CUBIC, borderMode=cv2.BORDER_CONSTANT)

If this is correct, then I should be able to find the bounding rectangle again and it should be (0, 0, bwidth, bheight):

corners = np.array([
  [0, 0],
  [0, height - 1],
  [width - 1, height - 1],
  [width - 1, 0]
])
corners = cv2.perspectiveTransform(np.float32([corners]), pth)[0]    
bx2, by2, bwidth2, bheight2 = cv2.boundingRect(corners)

print bx, by, bwidth, bheight
print bx2, by2, bwidth2, bheight2

Instead I get

-229 -82 947 1270
265 134 671 1096

Ugh due to more stupid karma rules I can't even answer my own question.

Matt Freeman wrote,

Homographies can be combined using matrix multiplication (which is why they are so powerful). If A and B are homographies, then AB represents the homography ...