Find point correspondences without features or descriptors?

asked 2020-10-18 06:50:45 -0600

antithing gravatar image

I have a set of data that is as follows:

std::vector<cv::Point3d> points3d; //3d points in world space
std::vector<cv::Point2d> points2d; //2d points in a camera view

cv::Mat intrinsics; // camera intrinsics
cv::Mat dist; //distortion parameters.

This equates to some white circles on a black wall (which I have the 3d positions of), and some camera frames of the markers/wall.

What I want to do is use a PnP solve to localize the camera. however, due to the lack of image features, I cannot use feature matching or descriptors, as I normally would.

How can I calculate correspondences in this case? I was thinking something like:

For each three markers in camera view, run p3p, return a pose, reproject the 3d points to 2d and measure the distance. repeat until the distance is small enough that we have the correct pose.

is there a better way? Can I use RANSAC to calculate the correspondences without giving it any prior matches? Can I use KNN somehow?

Thank you.

edit retag flag offensive close merge delete