kevinpaulose05 / ransac

This projects provides practical experience with essential matrix estimation, disambiguating relative pose, triangulation and analyzing reprojection errors for multiview 3D reconstruction. This involves estimating the essential matrix between two views and using it to recover the relative camera pose and 3D structure. The essential matrix E encapsulates the epipolar constraints between matching points in the two views. E is estimated from point matches using least squares and then refined with RANSAC to remove outliers. The epipolar geometry is leveraged to draw epipolar lines and evaluate E.

Four pose candidates are obtained from the SVD of E and -E to resolve the inherent ambiguity. Each candidate pose is used to triangulate matches to 3D points. The best pose is selected by counting how many points lie in front of both cameras. This allows recovering the relative rotation R and translation T up to scale. With the pose and 3D points known, reprojection errors can be visualized between points and their projections from the other view. The key learning is using fundamental matrices to encapsulate multi-view constraints for robust estimation and 3D reconstruction.