My guess is they estimate one or several dominant scene planes from the sparse triangulated feature points and get scale through incorporating accelerometer mea

My guess is they estimate one or several dominant scene planes from the sparse triangulated feature points and get scale through incorporating accelerometer measurements.