Inconsistent orientation of detected/generated ArUco and AprilTags

[qian-chu]

ArUco and AprilTag detection both uses a clockwise orientation, however, AprilTag corners start with bottom right while ArUco starts with upper left. See the example below:

Code:

import cv2


def draw_markers_with_corners(img, corners):
    for i, marker_corners in enumerate(corners):
        # Get the corners of the marker
        corner_points = marker_corners.reshape((4, 2))
        corner_points = corner_points.astype(int)

        # Draw each corner with its ordinal position
        for j, corner in enumerate(corner_points):
            # Draw corner point
            cv2.circle(img, tuple(corner), 5, (0, 0, 255), -1)
            # Label the corner with its ordinal position (0, 1, 2, 3)
            cv2.putText(
                img,
                str(j),
                (corner[0] + 10, corner[1] + 10),
                cv2.FONT_HERSHEY_SIMPLEX,
                0.5,
                (255, 0, 0),
                2,
            )

    return img


img = cv2.imread("markers.jpeg")

parameters = cv2.aruco.DetectorParameters()

april_dict = cv2.aruco.getPredefinedDictionary(cv2.aruco.DICT_APRILTAG_36h11)
april_detector = cv2.aruco.ArucoDetector(april_dict, parameters)
aruco_dict = cv2.aruco.getPredefinedDictionary(cv2.aruco.DICT_4X4_250)
aruco_detector = cv2.aruco.ArucoDetector(aruco_dict, parameters)

img_marked = img.copy()
april_corners, _, _ = april_detector.detectMarkers(img)
img_marked = draw_markers_with_corners(img_marked, april_corners)

aruco_corners, _, _ = aruco_detector.detectMarkers(img)
img_marked = draw_markers_with_corners(img_marked, aruco_corners)

cv2.imwrite("markers_detected.jpeg", img_marked)

The same issue persists (likely due to pre-defined dictionary configurations) for generation:

The generated AprilTag 36h11 - 0 (wrong orientation, see above)

The generated ArUco 4x4_250 - 0 (correct orientation, see above)

Code:

import cv2

april_dict = cv2.aruco.getPredefinedDictionary(cv2.aruco.DICT_APRILTAG_36h11)
aruco_dict = cv2.aruco.getPredefinedDictionary(cv2.aruco.DICT_4X4_250)

img = cv2.aruco.generateImageMarker(aruco_dict, 0, 200, borderBits=1)
cv2.imwrite("aruco.png", img)

img = cv2.aruco.generateImageMarker(april_dict, 0, 200, borderBits=1)
cv2.imwrite("april.png", img)

[qian-chu]

Also reported in https://stackoverflow.com/questions/79044142/why-is-the-order-of-the-incoming-corners-different-between-apriltag-and-aruco-ma

[s-trinh]

The generated image uses: https://github.com/opencv/opencv/blob/55a8996ffe660444b0a3c268c967f67b133ab332/modules/objdetect/src/aruco/apriltag/predefined_dictionaries_apriltag.hpp#L11961-L11966

{ {33, 161, 70, 186, 11},

33 in decimal gives in binary: 0010 0001
161 in decimal gives in binary: 1010 0001

---

In my opinion, the following should be tested/verified:

• what is the convention about the corners order in the official AprilTag repo?
• does the current output/convention match the AprilTag repo?
• what is the consequence about the estimated tag pose?
• code breaking change? add a parameter for that?

[s-trinh]

AprilTag official repository

Using the opencv_demo.cc in the AprilTag repository, it gives the following corners order:

You can use this tool to generate the marker images:

• https://chev.me/arucogen/

Regeneration of the AprilTag codeword

I have regenerated the AprilTag codeword in order to try to match the ArUco convention:

• https://github.com/s-trinh/opencv/blob/cb6a21ac3dc436dbeb8e4901e638ec1c4c3d91d2/modules/objdetect/src/aruco/apriltag/predefined_dictionaries_apriltag.hpp

It seems to work.

25h9

36h11

Tag image generation

Using your Python script and with the following parameters, it gives for:

• 36h11, id=46