Reprojection constraints can skew the optimization by a lot. The goal of this is to normalize each images reproejction constraints to be in line with what would be its odometry constraint (relative pose)
compute shannon entropy (H) from the pose covariance
Divide H by # of features in the image to get H per reprojection constraint
Reprojection constraints can skew the optimization by a lot. The goal of this is to normalize each images reproejction constraints to be in line with what would be its odometry constraint (relative pose)
May want to switch to dense_schur for the solver