From the essential matrix decomposition, we obtain rotation and translation in pairs of reflected matrices. This leads to 4 possible (Rc,c', tc,c') pairs for each feature triangulation.
I'm using the essential matrix estimation within RANSAC to detect outliers. So, in each Ransac iteration, I do the following:
- Estimate essential matrix using the fivepoint algorithm.
- Decompose the essential matrix leading to 4 possible
(Rc,c', tc,c'). - Compute the number of inliers correspoding to a reprojection error below a precalculated threshold for each 3D point of the dataset.
My question is, should I repeat the calculation of reprojection error for the 4 possibilities of the 3D point or it is possible to perform disambiguation and obtain only one 3D point before the computation of the reprojection error for the whole dataset?
In case it is possible to perform disambiguation beforehand, we will need 3 3D points, how to ensure that these points are inliers and not ouliers?
