我们有一个 ELP 1.0 百万像素双镜头 USB 立体相机，我们正在尝试使用 C++ 中的 OpenCV 3.1 来校准它。然而，校准的结果完全无法使用，因为调用stereoRectify完全扭曲了图像。这就是我们所做的：
在两个相机中找到校准（棋盘）图案，棋盘尺寸为 5x7，无论拍摄图像数量多少，结果几乎相同
findChessboardCorners(img[k], boardSize, corners, CALIB_CB_ADAPTIVE_THRESH | CALIB_CB_NORMALIZE_IMAGE)
cornerSubPix(img[k], corners, Size(11, 11), Size(-1, -1), TermCriteria(TermCriteria::COUNT+TermCriteria::EPS, 30, 0.01));
所有棋盘均已正确检测并使用验证
drawChessboardCorners(img[k], boardSize, corners, bFound);
然后我们分别对每个摄像头进行校准（不过这一步对于立体校准来说似乎并不重要），但是我们可以用它来分别验证每个摄像头
calibrateCamera(objectPoints, imagePoints[k], Size(320, 240), cameraMatrix[k], distCoeffs[k], rvecs, tvecs, 0)
然后我们进行立体校准
stereoCalibrate(objectPoints, imagePoints[0], imagePoints[1], cameraMatrix[0], distCoeffs[0], cameraMatrix[1], distCoeffs[1],
Size(320, 240), R, T, E, F, CALIB_USE_INTRINSIC_GUESS);
计算整流变换
stereoRectify(cameraMatrix[0], distCoeffs[0], cameraMatrix[1], distCoeffs[1], Size(320, 240), R, T, R1, R2, P1, P2, Q,
CALIB_ZERO_DISPARITY, 1, Size(320, 240), &validRoI[0], &validRoI[1]);
初始化映射以进行重新映射
Mat rmap[2][2];
initUndistortRectifyMap(cameraMatrix[0], distCoeffs[0], R1, P1, Size(FRAME_WIDTH, FRAME_HEIGHT), CV_16SC2, rmap[0][0], rmap[0][1]);
initUndistortRectifyMap(cameraMatrix[1], distCoeffs[1], R2, P2, Size(FRAME_WIDTH, FRAME_HEIGHT), CV_16SC2, rmap[1][0], rmap[1][1]);
...
remap(img, rimg, rmap[k][0], rmap[k][1], INTER_LINEAR);
imshow("Canvas", rimg);
结果是图像完全扭曲。正如我在一开始所说的，所有校准/棋盘图案都被正确检测到，如果我们不调用stereoRectify函数，未失真的图像（重新映射后）看起来很完美。如果我们调用stereoRectify函数就会出现问题。
我们是否错过了什么？ 校准图像的数量似乎没有任何影响（有时拍摄 2 张图像比 10 张图像提供更好的结果（但仍然不可用））
这是校准图案的示例。我们采取几个不同的方向：
This is the result of the calibration if we do not call stereoRectify:
This is the wrong result if we call stereoRectify (but mostly it gets much worse):
预先感谢您提供任何可能出现问题的帮助。
只是总结一下，如果有人需要类似的帮助，这就是我为获得最佳外观结果所做的事情：
在角点检测之前放大棋盘图像：
Mat resized;
resize(img[k], resized, Size(FRAME_WIDTH * 2, FRAME_HEIGHT * 2), 0.0, 0.0, INTER_LINEAR);
findChessboardCorners(resized, boardSize, corners, CALIB_CB_ADAPTIVE_THRESH | CALIB_CB_NORMALIZE_IMAGE
缩小检测到的角点：
for (int i = 0; i < corners.size(); ++i) {
corners[i].x /= 2.0;
corners[i].y /= 2.0;
}
分别校准每个相机：
double rms = calibrateCamera(objectPoints, imagePoints[k], Size(FRAME_WIDTH, FRAME_HEIGHT), cameraMatrix[k], distCoeffs[k], rvecs, tvecs,
CALIB_FIX_PRINCIPAL_POINT | CALIB_FIX_ASPECT_RATIO | CALIB_ZERO_TANGENT_DIST | CALIB_RATIONAL_MODEL | CALIB_FIX_K3 | CALIB_FIX_K4 | CALIB_FIX_K5);
校准立体相机：
stereoCalibrate(objectPoints, imagePoints[0], imagePoints[1], cameraMatrix[0], distCoeffs[0], cameraMatrix[1], distCoeffs[1],
Size(FRAME_WIDTH, FRAME_HEIGHT), R, T, E, F,
CALIB_FIX_INTRINSIC | CALIB_SAME_FOCAL_LENGTH,
TermCriteria(TermCriteria::COUNT + TermCriteria::EPS, 30, 0));
计算校正（alpha = 0.0）：
stereoRectify(cameraMatrix[0], distCoeffs[0], cameraMatrix[1], distCoeffs[1], Size(FRAME_WIDTH, FRAME_HEIGHT),
R, T, R1, R2, P1, P2, Q,
CALIB_ZERO_DISPARITY, 0.0, Size(FRAME_WIDTH, FRAME_HEIGHT), &validRoI[0], &validRoI[1]);
这些是校准结果矩阵
内在：
M1: !!opencv-matrix
rows: 3
cols: 3
dt: d
data: [ 2.6187262304487734e+02, 0., 1.5950000000000000e+02, 0.,
2.6187262304487734e+02, 1.1950000000000000e+02, 0., 0., 1. ]
D1: !!opencv-matrix
rows: 1
cols: 5
dt: d
data: [ -4.6768074176991381e-01, 2.0221327568191746e-01, 0., 0., 0. ]
M2: !!opencv-matrix
rows: 3
cols: 3
dt: d
data: [ 2.6400975025525213e+02, 0., 1.5950000000000000e+02, 0.,
2.6400975025525213e+02, 1.1950000000000000e+02, 0., 0., 1. ]
D2: !!opencv-matrix
rows: 1
cols: 5
dt: d
data: [ -4.5713211677198845e-01, 2.8855737500717565e-01, 0., 0., 0. ]
外在因素：
R: !!opencv-matrix
rows: 3
cols: 3
dt: d
data: [ 9.9963073433190641e-01, 4.6310793035473068e-04,
2.7169477545556639e-02, -6.9475632716349024e-04,
9.9996348636555088e-01, 8.5172324905818230e-03,
-2.7164541091274301e-02, -8.5329635354663789e-03,
9.9959455592785362e-01 ]
T: !!opencv-matrix
rows: 3
cols: 1
dt: d
data: [ -6.1830090720273198e+01, 1.6774590574449604e+00,
1.8118983433925613e+00 ]
我的另一个问题是变量初始化是否有任何特殊要求，或者这是否足够？
Mat cameraMatrix[2] = { Mat::eye(3, 3, CV_64F), Mat::eye(3, 3, CV_64F) };
Mat distCoeffs[2], R, T, E, F, R1, R2, P1, P2, Q;