使用Raspberry Pi相机模块时，cv2 findChessboardCorner精度较低

Question

我正在尝试校准两台摄像机以获得立体视觉。一个是Pi Camera，另一个是插入Raspberry Pi的USB网络摄像头。我不知道为什么，但是Pi发现的棋盘点准确性（总是）不高。请看所附图片。合适的摄像头（网络摄像头）更简单，更准确。有什么办法可以改善这一点？

这是脚本的某些部分。

# webcam is called by cv2.VideoCapture, and picam is called by PiCamera()
left = PiCamera()
left.resolution = (640,480)
right = cv2.VideoCapture(0)

while n <= Total_num:

    right.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
    right.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)

    img1 = np.empty((480, 640, 3), dtype=np.uint8)
    left.capture(img1, "bgr")
    ret, img2 = right.read()

此后，它们变成了灰色，并且找到了角落。

gray1 = cv2.cvtColor(img1, cv2.COLOR_BGR2GRAY)
gray2 = cv2.cvtColor(img2, cv2.COLOR_BGR2GRAY)

ret1, corners1 = cv2.findChessboardCorners(gray1, checkerBoard, None)
ret2, corners2 = cv2.findChessboardCorners(gray2, checkerBoard, None)

cv2.cornerSubPix(gray1, corners1, (5, 5), (-1, -1), criteria)
cv2.cornerSubPix(gray2, corners1, (5, 5), (-1, -1), criteria)

cv2.drawChessboardCorners(img1, checkerBoard, corners1, ret1)
cv2.drawChessboardCorners(img2, checkerBoard, corners2, ret2)