Author: Wang
Year: 2020
Notes:
- here the dense prior on robot pose is kept, while the dense prior on landmark only is sparsified: recover only landmark-to-pose pseudo factors
- marginalization can cause inacuraccy as the linearization point of the prior is fixed
- when a KF is marginalized:
- the pose, velocity and biases are marginalized into a dense prior H_d as in OKVIS & co
- the landmark only connected to the KF are marginalized into another dense prior H_o that will be partially sparsified
- Uses NFR to sparsify H_o (is J really invertible?)
Results:
- Way faster than MKaess solution