cv2.calibrateHandEye 在提供 3x1 旋转向量或 3x3 旋转矩阵时输出不同的结果

Question

我目前正在尝试使用 opencv 的手眼校准功能 (cv2.calibrateHandEye()) 来校准 Microsoft Azure Kinect。但是，当我插入 3x1 旋转向量而不是 3x3 旋转矩阵作为输入时，我使用所有方法得到完全不同的结果。下面，我将分享两种情况下的输出结果。

我想知道为什么会出现这种差异，因为我使用 cv2.Rodrigues 在向量和矩阵之间进行转换。

3x1 Rotation Vector case:
(19, 3) (19, 3) (19, 3) (19, 3)       
--------------------------------------
Method 0
Rotation:
 [[1. 0. 0.]
 [0. 1. 0.]
 [0. 0. 1.]]
Translation:
 [[0.]
 [0.]
 [0.]]
--------------------------------------
--------------------------------------
Method 1
Rotation:
 [[nan nan nan]
 [nan nan nan]
 [nan nan nan]]
Translation:
 [[0.]
 [0.]
 [0.]]
--------------------------------------
--------------------------------------
Method 2
Rotation:
 [[-1.  0.  0.]
 [ 0. -1.  0.]
 [ 0.  0.  1.]]
Translation:
 [[0.]
 [0.]
 [0.]]
--------------------------------------
--------------------------------------
Method 3
Rotation:
 [[-0.03388477  0.81429681  0.57945882]
 [ 0.37003643 -0.52836555  0.76413538]
Translation:
 [[0.]
 [0.]
 [0.]]
--------------------------------------
--------------------------------------
Method 4
Rotation:
 [[nan nan nan]
 [nan nan nan]
 [nan nan nan]]
Translation:
 [[nan]
 [nan]
 [nan]]
--------------------------------------

3x3 Rotation Matrix case:
(19, 3, 3) (19, 3) (19, 3, 3) (19, 3)
--------------------------------------
Method 0
Rotation:
 [[ 0.19681749  0.15272093 -0.96847261]
 [-0.9802481   0.01110192 -0.19745987]
 [-0.01940435  0.988207    0.15188945]]
Translation:
 [[426.01991564]
 [  6.31112392]
 [212.62483639]]
--------------------------------------
--------------------------------------
Method 1
Rotation:
 [[ 4.38898532e-04  7.05825236e-02  9.97505847e-01]
 [-9.99993453e-01 -3.55184483e-03  6.91318082e-04]
 [-3.59178096e-03  9.97499620e-01 -7.05805026e-02]]
Translation:
 [[ 113.83854629]
 [ -64.48053741]
 [-155.89394605]]
--------------------------------------
--------------------------------------
Method 2
Rotation:
 [[-1.63313542e-04  1.12303599e-01 -9.93673928e-01]
 [-9.99994011e-01 -3.45353893e-03 -2.25961739e-04]
 [-3.45706791e-03  9.93667940e-01  1.12303491e-01]]
Translation:
 [[429.63163945]
 [-72.98653944]
 [228.31050502]]
--------------------------------------
--------------------------------------
Method 3
Rotation:
 [[ 0.27525174  0.87240146  0.403921  ]
 [ 0.20267468  0.35804992 -0.91144019]
 [-0.93976564  0.33274006 -0.07825982]]
Translation:
 [[-126.85624931]
 [ 205.9095892 ]
 [  90.09963159]]
--------------------------------------
--------------------------------------
Method 4
Rotation:
 [[-0.20970054  0.33923475 -0.91703079]
 [-0.94475058 -0.31195916  0.1006371 ]
 [-0.25193655  0.88746902  0.38591023]]
Translation:
 [[ 716.47999662]
 [-112.67040497]
 [ 604.15236449]]
--------------------------------------

Answer 1

对你来说可能为时已晚，但对其他人来说可能很有趣：我偶然发现了同样的错误，在我的例子中，我有一个用于 R_gripper2base 的 Mat 向量。我使用 push_back 将实际旋转放在列表的末尾。当我将向量作为输入时，一切正常，但使用罗德里格斯矩阵 push_back 不仅将实际矩阵放在末尾，而且还更改了列表中每个其他条目的值。所以最后我得到了一个完全相同的矩阵向量。这就是为什么 calibrateHandEye 的结果很糟糕。

cv2.calibrateHandEye 在提供 3x1 旋转向量或 3x3 旋转矩阵时输出不同的结果

cv2.calibrateHandEye outputs different results when provided 3x1 rotation vector or 3x3 rotation matrix

opencv

calibration

cv2