Nonparametric Segment Detection

Abstract

In computer and robotic vision point clouds from depth sensors
have to be processed to form higher-level concepts such as lines,
planes, and objects. Bayesian methods formulate precisely prior knowledge
with respect to the noise and likelihood of points given a line, plane,
or object. Nonparametric methods also formulate a prior with respect
to the number of those lines, planes, or objects. Recently, a nonparametric
Bayesian method has been proposed to perform optimal inference
simultaneously over line fitting and the number of lines. In this paper
we propose a nonparametric Bayesian method for segment fitting. Segments
are lines of finite length. This requires 1.) a prior for line segment
lengths: the symmetric Pareto distribution, 2.) a sampling method that
handles nonconjugacy: an auxiliary variable MCMC method. Results
are measured according to clustering performance indicators, such as
the Rand Index, the Adjusted Rand Index, and the Hubert metric. Surprisingly,
the performance of segment recognition is worse than that of
line recognition. The paper therefore concludes with recommendations
towards improving Bayesian segment recognition in future wo