Initial Testing

To move from simulated data to a collected pose-graph, the manhattan3500 data-set was utilized – this pose-graph is depicted in Figure 1. Initially, we process the fault-free pose-graph and the pose-graph that has 100 false constraints added.

Fig 1 :: Manhattan 3500 pose-graph

After processing both pose-graphs, the residuals from each graph can be analyzed. The residual scatter plots for the fault-free graph and the faulty graph are shown in Figures 2 and 3, respectively. This plots depict what would be expected; both distributions have approximately zero mean with the faulty graph residuals have a larger variance.

Fig 2 :: Fault free residual scatter

Fig 3 :: Pose graph residual scatter when 100 faulty constraints are added

Using the faulty residuals, the BMM DP was tested. To allow for easy visualization, the classification is shown in 2D slices of the 3D residual scatter (i.e., Figure 4 only shows the classification of the X-Y subset of the data).

Fig 4 :: Classification on X-Y residuals

Fig 5 :: Classification on X-Heading residuals

Fig 6 :: Classification on Y-Heading residuals

From the three figures above, it can be seen that the BMM DP is categorizing the data into two bins ( i.e., inlier or outlier ) without being instructed that there are two categories present in the data-set.

Next Steps ::

Next, the mixture components provided by the BMM DP algorithm, which are shown below, will be incorporated into the max-mixtures framework to see how well this Gaussian mixture preforms.