Robust Estimation in High Noise and Highly Dimensional Data Sets with Applications to Machine Vision
Darren R. Myatt · 2002
Robust estimation, the problem of providing an effective parameter estimation in the presence of outlying data, is central to the discipline of machine vision. This report examines current robust estimation methods and demonstrates their inadequacy in high noise and high dimensionality due to reliance on naive hypothesis selection. Subsequently, clustering algorithms are discussed and how ideas from this area may be introduced into modern robust estimators to increase their efficacy is hypothesised. A preliminary algorithm using enhanced data sampling, NAPSAC, is described and shown to be superior in both high noise and higher dimensions. Based on this success, possibilities for further work are introduced along with some conclusions. 2.4 Generic random sampling algorithm............... 21