Our contributions

1
Edwin Olson, MITMatthew Walter, MITJohn
Leonard, MITSeth Teller, MIT http//cgr.csail.mit
.edu
Single Cluster Graph Partitioningfor Robotics
Applications

• Our contributions
• A spectral clustering algorithm that can decide
• This point isnt part of any cluster
• Useful for real-world data sets with noise and
  mistakes
• We extend clustering to non-square/non-symmetric
  matrices
• Prior work
• Ncuts, MinMax cuts
• No outlier handling assumes all points belong to
  a logical cluster
• Basic Idea
• Find a set of points which are maximally mutually
  consistent.

NCUT
SCGP
2
Edwin Olson, MITMatthew Walter, MITJohn
Leonard, MITSeth Teller, MIT http//cgr.csail.mit
.edu
Single Cluster Graph Partitioningfor Robotics
Applications

• Outlier Rejection
• Inputs Noisy measurements
• Outputs Inliers
• Data Association
• Inputs data-association hypotheses
• Outputs self-consistent data-association
  hypotheses
• Feature Detection/Estimation
• Inputs points and proposed models
• Outputs self-consistent points and models
• Lower error and faster than RANSAC

Multi-track sonar range data
No Outlier Rejection
Outlier Rejection (SCGP)
Observations
Map
Data Associations
Line detection and extraction