Because it is not stereo data, the variance computation should divide only by 2. Currently the linear variance would be underestimated, causing some upcoming good loop closures to be rejected (by loop closure links, not odometry links).