OpenCV Java版projectPoints函数基础示例疑似异常

Hey there! Let's dig into why your projectPoints call isn't matching your expected results. Since you mentioned you have a working undistorted implementation, let's walk through the most common missing steps or misconfigurations that could cause this discrepancy:

• Verify Camera Intrinsics Matrix (K) Format
  Make sure your camera matrix follows the standard layout expected by the function (e.g., OpenCV's projectPoints uses [fx, 0, cx; 0, fy, cy; 0, 0, 1]). Your expected u-values are all 0.0, which implies your principal point x-coordinate (cx) should be 0.0, and focal length fx should align such that (fx * x)/z + cx = 0 for your 3D points. A non-zero cx or incorrect fx will immediately throw off your u-values.

• Check 3D Input Point Structure
  Confirm your 3D points use non-homogeneous (x,y,z) coordinates (not homogeneous (x,y,z,w)). For your expected v-values (100.0, 20.0, 10.0), your 3D points should have a consistent z-value (or y/z ratio) that matches your focal length fy and principal point y-coordinate cy. For example, if fy=1 and cy=0, your 3D points' y-values should directly match the expected v-values when z=1.

• Ensure Distortion Coefficients Are Zero
  Since your custom implementation is undistorted, double-check that you're passing a zero-filled distortion coefficients array to projectPoints. Even tiny non-zero values here will warp your output points away from the expected undistorted results. For the standard 5-parameter distortion model, this means passing [0, 0, 0, 0, 0].

• Confirm Rotation/Translation Vectors Are Neutral
  If your scene's world coordinate system aligns with the camera's coordinate system, your rotation vector (rvec) and translation vector (tvec) must be set to zero. Any non-zero rotation will rotate your 3D points, breaking the u=0.0 pattern, while non-zero translation will shift your projected points off their expected positions.

• Avoid Double-Applying Intrinsics
  Some projectPoints implementations automatically apply the camera intrinsics matrix to convert camera-coordinate points to pixel coordinates. If you're manually multiplying the output by your K matrix again, you'll get incorrect scaled/shifted results. Compare this to your custom implementation—if yours applies K once, make sure the library function isn't doing it for you (or vice versa).

Example Correct Call (Matching Your Expected Output) #

Here's how a properly configured call might look (using OpenCV's Java API as an example):

// Set up neutral extrinsic parameters (camera = world origin)
Mat rvec = Mat.zeros(3, 1, CvType.CV_64F);
Mat tvec = Mat.zeros(3, 1, CvType.CV_64F);

// Zero distortion coefficients for undistorted projection
Mat distCoeffs = Mat.zeros(5, 1, CvType.CV_64F);

// Camera matrix matching your expected output (fx=1, fy=1, cx=0, cy=0)
Mat K = new Mat(3, 3, CvType.CV_64F);
K.put(0, 0, 1.0);
K.put(0, 2, 0.0);
K.put(1, 1, 1.0);
K.put(1, 2, 0.0);
K.put(2, 2, 1.0);

// 3D points that project to your expected (u,v) values
List<Point3d> objectPoints = Arrays.asList(
    new Point3d(0.0, 100.0, 1.0),
    new Point3d(0.0, 20.0, 1.0),
    new Point3d(0.0, 10.0, 1.0)
);

// Run projection
Mat imagePoints = new Mat();
Calib3d.projectPoints(
    Converters.vector_Point3d_to_Mat(objectPoints),
    rvec, tvec, K, distCoeffs, imagePoints
);

// Convert output to your preferred format
List<Point2d> result = Converters.Mat_to_vector_Point2d(imagePoints);
// Result should match your expected [{0.0, 100.0}, {0.0, 20.0}, {0.0, 10.0}]

Start by cross-checking each of these points against your current code—odds are one of these configurations is out of sync with your custom undistorted implementation.

内容的提问来源于stack exchange，提问作者J.E.Tkaczyk