Cropping to only valid pixels after remap with fisheye (equidistant) stereo rectification?

Question

I have two cameras offset horizontally and have acquired their calibration parameters (Camera Matrix and Distortion Coefficients, as well as the transform between them) using Kalibr under the pinhole-equidistant model (distortion coefficients k1, k2, k3, k4).

I want to use openCVs cv.fisheye.stereoRectify to create new projection matrices for each camera that i can feed into cv.fisheye.initUndistortRectifyMap and then into cv.remap to rectify and undistort each image.

Unfortunately even with the balance parameter in fisheye.stereoRectify set to 0 the remaped images still have black pixels bowing into them. I want to crop each image such that no invalid pixels exist in either of the undistorted camera images.

I see that the standard cv.stereoRectify function has an alpha parameter that does exactly this. But it seems like cv.fisheye.stereoRectify does not have this parameter. Thus I want to replicate its features.

cv.stereoRectify seems to use the radtan distortion model (distortion parameters k1, k2, p1, p2) so I dont think i can easily swap that function in since i dont have p1 and p2.

snippet from my pipeline to follow below:

R1, R2, P1, P2, Q = cv2.fisheye.stereoRectify(mtx_right, dist_right, 
                                              mtx_left, dist_left, 
                                              (960,1280), R, tvec, 
                                              flags=cv2.CALIB_ZERO_DISPARITY,
                                              balance= 0.0, fov_scale=1)

map1_right, map2_right = cv2.fisheye.initUndistortRectifyMap(mtx_right, dist_right, 
                                                             R1, P1[0:3, 0:3], 
                                                             (1280, 960), cv2.CV_16SC2)  

map1_left, map2_left = cv2.fisheye.initUndistortRectifyMap(mtx_left, dist_left,
                                                           R2, P2[0:3, 0:3],
                                                           (1280, 960), cv2.CV_16SC2)

undistorted_right = cv2.remap(img_rgb_right, map1_right, map2_right, 
                             interpolation=cv2.INTER_LINEAR,
                             borderMode=cv2.BORDER_CONSTANT)
undistorted_left = cv2.remap(img_rgb_left, map1_left, map2_left, 
                            interpolation=cv2.INTER_LINEAR,
                            borderMode=cv2.BORDER_CONSTANT)

Is there an easy way to get the same functionality that alpha produces in the traditional cv.stereoRectify? balance=0 seems close but doesn't completely cut off the invalid pixels.

CURRENT OUTPUT (balance=0.5 to zoom out a little)

GOAL IS FOR BOTH IMAGES TO ONLY SHOW WHATS IN THE GREEN BOX (same dimensions on either if that isn't clear,whichever has a smaller valid pixel rectangle):

Answer 1

The right way to do it is to simply set the alpha value to 0.0. In the OpenCV documentation: alpha=0 means that the rectified images are zoomed and shifted so that only valid pixels are visible (no black areas after rectification).

---

Alternatively, you can implement your own algorithm for this, something like:

Imagine that I have these two images, after rectifying (the green lines are just to visually check that the rectification is correct):

R1, R2, P1, P2, Q, roi1, roi2 = \
        cv2.stereoRectify(cameraMatrix1=k_1,
                          distCoeffs1=dist_coeff,
                          cameraMatrix2=k_2,
                          distCoeffs2=dist_coeff,
                          imageSize=(width, height),
                          R=r_stereo,
                          T=t_stereo,
                          flags=cv2.CALIB_ZERO_DISPARITY,
                          alpha=1.0
                          )

map1x, map1y = cv2.initUndistortRectifyMap(
    cameraMatrix=k_1,
    distCoeffs=dist_coeff,
    R=R1,
    newCameraMatrix=P1,
    size=(width, height),
    m1type=cv2.CV_32FC1)

map2x, map2y = cv2.initUndistortRectifyMap(
    cameraMatrix=k_2,
    distCoeffs=dist_coeff,
    R=R2,
    newCameraMatrix=P2,
    size=(width, height),
    m1type=cv2.CV_32FC1)

im_1_rect = cv2.remap(im_1, map1x, map1y, cv2.INTER_LINEAR)
im_2_rect = cv2.remap(im_2, map2x, map2y, cv2.INTER_LINEAR)

result = np.hstack((im_1_rect, im_2_rect))
for tmp_col in range(20, height, 30):
    result = cv2.line(result, (0, tmp_col), (int(2.0 * width), tmp_col), (0, 255, 0), 1)

cv2.imshow("rectified image", result)
cv2.waitKey(0)

The trick is to project the points in the border of the image to the rectified image and then check the projected u and v coordinates. For example, here I will show you how we can project the corners of the original image to the rectified images (painted in red):

pts = np.array([[[0, 0]], [[width - 1, 0]], [[0, height - 1]], [[width - 1, height - 1]]], dtype=np.float64)
pts_transformed_l = cv2.undistortPoints(pts, k_1, dist_coeff, R=R1, P=P1)
pts_transformed_r = cv2.undistortPoints(pts, k_2, dist_coeff, R=R2, P=P2)
for pt in pts_transformed_l:
    u, v = pt[0]
    result = cv2.circle(result, (int(round(u)), int(round(v))), 3, (0, 0, 255), -1)
for pt in pts_transformed_r:
    u, v = pt[0]
    u += 640
    result = cv2.circle(result, (int(round(u)), int(round(v))), 3, (0, 0, 255), -1)
cv2.imshow("rectified image with corners", result)
cv2.waitKey(0)

Let's start with the cropping the height. To do that, we warp the top and bottom points to each of the rectified images. We want the largest v for the top part and the smallest v for the bottom part. Note that now we need to consider all the points of the top (0, 0), (0, 1), ... (width-1, 0) since the rectified images have generally curved borders. In other words, we basically want to find the largest v in the red points and the smallest v in the points in blue in this image:

""" crop in the v direction """
pts_top_list = []
pts_bot_list = []
for i in range(width):
    pt_tmp = [[i, 0]]
    pts_top_list.append(pt_tmp)
    pt_tmp = [[i, height]]
    pts_bot_list.append(pt_tmp)
pts_top = np.asarray(pts_top_list, dtype=np.float64)
pts_bot = np.asarray(pts_bot_list, dtype=np.float64)

# top part - larger v
v_top = 0
## rectified image 1
pts_transformed_l = cv2.undistortPoints(pts_top, k_1, dist_coeff, R=R1, P=P1)
for pt in pts_transformed_l:
    _, v = pt[0]
    if math.ceil(v) > v_top:
        v_top = math.ceil(v)
## rectified image 2
pts_transformed_r = cv2.undistortPoints(pts_top, k_2, dist_coeff, R=R2, P=P2)
for pt in pts_transformed_r:
    _, v = pt[0]
    if math.ceil(v) > v_top:
        v_top = math.ceil(v)

# bottom part - smaller v
v_bot = height
## rectified image 1
pts_transformed_l = cv2.undistortPoints(pts_bot, k_1, dist_coeff, R=R1, P=P1)
for pt in pts_transformed_l:
    _, v = pt[0]
    if int(v) < v_bot:
        v_bot = int(v)
## rectified image 2
pts_transformed_r = cv2.undistortPoints(pts_bot, k_2, dist_coeff, R=R2, P=P2)
for pt in pts_transformed_r:
    _, v = pt[0]
    if int(v) < v_bot:
        v_bot = int(v) 

result_cropped_v = result[v_top:v_bot, :]
cv2.imshow("rectified cropped v", result_cropped_v)
cv2.waitKey(0)

You can apply the same in the u direction. Just be careful that if you crop in u and you are estimating disparity, you will need to take that into consideration before estimating depth!

""" crop in the u direction (for both images) """
pts_left_list = []
pts_rght_list = []
for i in range(width):
    pt_tmp = [[0, i]]
    pts_left_list.append(pt_tmp)
    pt_tmp = [[width, i]]
    pts_rght_list.append(pt_tmp)
pts_left = np.asarray(pts_left_list, dtype=np.float64)
pts_rght = np.asarray(pts_rght_list, dtype=np.float64)

# rectified image 1
## left part - larger u
u_left_1 = 0
pts_transformed_l = cv2.undistortPoints(pts_left, k_1, dist_coeff, R=R1, P=P1)
for pt in pts_transformed_l:
    u, _ = pt[0]
    if math.ceil(u) > u_left_1:
        u_left_1 = math.ceil(u)
## right part - smaller u
u_right_1 = width
pts_transformed_r = cv2.undistortPoints(pts_rght, k_1, dist_coeff, R=R1, P=P1)
for pt in pts_transformed_r:
    u, _ = pt[0]
    if int(u) < u_right_1:
        u_right_1 = int(u)

# rectified image 2
## left part - larger u
u_left_2 = 0
pts_transformed_l = cv2.undistortPoints(pts_left, k_2, dist_coeff, R=R2, P=P2)
for pt in pts_transformed_l:
    u, _ = pt[0]
    if math.ceil(u) > u_left_2:
        u_left_2 = math.ceil(u)
## right part - smaller u
u_right_2 = width
pts_transformed_r = cv2.undistortPoints(pts_rght, k_2, dist_coeff, R=R2, P=P2)
for pt in pts_transformed_r:
    u, _ = pt[0]
    if int(u) < u_right_2:
        u_right_2 = int(u)

im_1_rect_cropped = im_1_rect[v_top:v_bot, u_left_1:u_right_1]
im_2_rect_cropped = im_2_rect[v_top:v_bot, u_left_2:u_right_2]
result_cropped = np.hstack((im_1_rect_cropped, im_2_rect_cropped))
for tmp_col in range(20, height, 30):
    result = cv2.line(result_cropped, (0, tmp_col), (int(2.0 * width), tmp_col), (0, 255, 0), 1)
cv2.imshow("rectified image cropped", result_cropped)
cv2.waitKey(0)

Answer 2

Here is how I would do it in Imagemagick using -trim. I note that -trim can keep track of the offsets of the upper left corner after the trim relative to where it was before the trim (by leaving off +repage, which clears that geometry information). So I trim each image and have it keep track. Then I place the trimmed images in a black background separately and then append the two results side-by-side and then trimmed the black again.

Since the originals were not provided, I cut the images out of what was provided.

Left:

Right:

magick left.png -format "%wx%h" -write info: -fuzz 15% -trim \
-fuzz 5% -define trim:percent-background=0 \
-define trim:background-color=black -trim left_im_trim.png

magick right.png -format "%wx%h" -write info: -fuzz 15% -trim \
-fuzz 5% -define trim:percent-background=0 \
-define trim:background-color=black -trim right_im_trim.png


magick \
\( left_im_trim.png -set page "%wx+0+%Y" -background black -flatten \)  \
\( right_im_trim.png -set page "%wx+0+%Y" -background black -flatten \) \
-background black +append \
-define trim:percent-background=0 \
-define trim:background-color=black \
-trim +repage left_right_trim_append.png

Left Trimmed:

Right Trimmed:

Appended and Trimmed Again:

I left the 3 commands above separate so that one could see the results. But they could all be combined into 1 long command line.