In the last update we looked into the sensitivity of the Dirichlet process to its hyper-parameter, $\alpha$. Today, we move to processing faulty pose-graphs with our modified max-mixtures model. To conduct the initial test, we will utilize the Manhattan 3500 pose graph, as depicted in Figure 1. Using this pose-graph, we add 50 false constraints to test the estimators ability to robustly optimize in the face of faulty measurements.



Figure 1 :: Manhattan 3500 pose graph



As a reference, we first optimized the faulty pose graph using $L_2$ optimization. This is depicted in the video provided below. As video depicts — and as expected — traditional $L_2$ optimization does not perform well when presented with a faulty graph.





Next, we optimize the graph using E.M. to estimate the covariance model and max-mixtures to select the most fitting component. When this model is employed, the optimized graph much more closely resembles what it would look like if no faults were present. A video showing the optimization is shown below.





The code used to conduct this test can be found here



Next Steps

Short term to-do list

  • Compare to other robust models
  • Test on several graphs
  • Generate confusion matrix to check classifier accuracy