OpenCV stereoCalibrateExtended报错：非平面标定板需指定初始内参

错误信息 #

cv2.error: OpenCV(4.10.0) /io/opencv/modules/calib3d/src/calibration.cpp:1682: error: (-5:Bad argument) For non-planar calibration rigs the initial intrinsic matrix must be specified in function 'cvCalibrateCamera2Internal'

用户代码 #

def stereo_calibrate(img, correspondences_left, correspondences_right, camera_matrix_L, camera_matrix_R, baseline=4.1, dist_coeffs=None):
    if dist_coeffs is None:
        dist_coeffs = np.zeros((4, 1))

    # Extract image points and object points from correspondences
    image_points_left = correspondences_left[:, :2].astype(np.float32)
    image_points_right = correspondences_right[:, :2].astype(np.float32)
    object_points = correspondences_left[:, 2:].astype(np.float32)  # Assuming object points are the same for both sets

    rvec_L, tvec_L = np.array([0., 0., 0.]), np.array([0, 0., 0.])
    rvec_R, tvec_R = np.array([-0.00637872, 0.00102092, 0.000673804]), np.array([0., 0., 0.])

    # Convert rotation vectors to rotation matrices
    R_L, _ = cv2.Rodrigues(rvec_L)
    R_R, _ = cv2.Rodrigues(rvec_R)
    print("after converting via Rodrigues:", R_L, R_R)

    # Initialize rotation and translation vectors
    R_init = R_R @ R_L.T
    # T_init = tvec_R - R_init @ tvec_L
    T_init = np.array([0., 0., 0.])
    print("R_init", R_init)
    print("T_init", T_init)

    # Stereo calibration to refine the extrinsic parameters
    ret, cameraMatrix1, distCoeffs1, cameraMatrix2, distCoeffs2, R, T, E, F, perViewErrors, _, _= cv2.stereoCalibrateExtended(
        [object_points], [image_points_left], [image_points_right],
        camera_matrix_L, dist_coeffs, camera_matrix_R, dist_coeffs,
        imageSize=(img.shape[1], img.shape[0]),  # Correct image size as (width, height)
        R=R_init, T=T_init,
        criteria=(cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 100, 1e-6),
        flags=cv2.CALIB_USE_EXTRINSIC_GUESS
    )
    print("ret", ret)
    if ret:
        print("Stereo Calibration Successful")
        print("Rotation Matrix between the cameras:")
        print(R)
        print("Translation Vector between the cameras:")
        print(T)
        print("dist_coeffs1", distCoeffs1)
        print("dist_coeffs2", distCoeffs2)
        print(rvec_L)
        print(tvec_L)
    else:
        raise ValueError("stereoCalibrate failed to find a solution")

    # Reproject object points to both cameras
    reprojected_points_left, _ = cv2.projectPoints(object_points, rvec_L, tvec_L, camera_matrix_L, distCoeffs1)
    reprojected_points_right, _ = cv2.projectPoints(object_points, R, T, camera_matrix_R, distCoeffs2)
    
    return R, T, reprojected_points_left, reprojected_points_right

correspondences_left = np.array([
    [671.0889955686854, 193.80354505169868, 2.4492935982947064e-16, -4.0, 11],
...
])

correspondences_right = np.array([
    [436.2245592329106, 193.79399938137954, 2.4492935982947064e-16, -4.0, 11],
...
])

# Camera intrinsic parameters, replace with your actual camera matrix
camera_matrix_L = np.array([
    [1018.9, 0, 601.447],
    [0, 1018.9, 517.462],
    [0, 0, 1]
], dtype=np.float32)

camera_matrix_R = np.array([
    [1018.9, 0, 690.392],
    [0, 1018.9, 517.462],
    [0, 0, 1]
], dtype=np.float32)

# calling the function
R, T, reprojected_points_left, reprojected_points_right = stereo_calibrate(img_left, correspondences_left, correspondences_right, camera_matrix_L, camera_matrix_R)

问题原因 #

1. flags参数未锁定内参：当前仅设置了CALIB_USE_EXTRINSIC_GUESS，允许OpenCV优化内参。但你的标定靶标是非平面的（物体点z坐标有变化），OpenCV在优化非平面靶标的标定时，需要明确的初始内参且不允许重新初始化，否则会触发该错误。
2. 物体点仅单帧数据：stereoCalibrateExtended需要多帧不同姿态的标定数据提供足够约束，单帧非平面点无法满足标定的自由度要求，也会导致内参初始化逻辑异常。
3. 畸变系数共用变量：左右相机使用同一个dist_coeffs变量，可能导致OpenCV在优化时出现参数冲突，间接影响内参的初始化判断。

解决方案 #

• 修改flags参数，锁定内参：添加CALIB_FIX_INTRINSIC，强制OpenCV使用你传入的初始内参，不进行优化：

  flags=cv2.CALIB_USE_EXTRINSIC_GUESS + cv2.CALIB_FIX_INTRINSIC
  

• 补充多帧标定数据：确保objectPoints、imagePoints1、imagePoints2是包含多帧数据的列表，例如：

  # 示例：假设有3帧数据
  object_points_list = [object_points_frame1, object_points_frame2, object_points_frame3]
  image_points_left_list = [img_points_left1, img_points_left2, img_points_left3]
  image_points_right_list = [img_points_right1, img_points_right2, img_points_right3]
  

  然后将这三个列表传入stereoCalibrateExtended。
• 单独初始化左右相机畸变系数：避免共用变量，分别初始化：

  dist_coeffs_L = np.zeros((4, 1), dtype=np.float32)
  dist_coeffs_R = np.zeros((4, 1), dtype=np.float32)
  

  并在函数调用时传入这两个变量：

  ret, cameraMatrix1, distCoeffs1, cameraMatrix2, distCoeffs2, R, T, E, F, perViewErrors, _, _= cv2.stereoCalibrateExtended(
      [object_points], [image_points_left], [image_points_right],
      camera_matrix_L, dist_coeffs_L, camera_matrix_R, dist_coeffs_R,
      # 其他参数不变
  )
  

内容的提问来源于stack exchange，提问作者Yajax