||Unsupervised segmentation evaluation measures are typically based on smoothness
within segments and contrast between them. Due to the relative simplicity of
these measures, optimizing them often leads to elegant segmentation algorithms.
However, the simplicity of the measures does not necessarily lead to good
segmentation. Moreover, as an evaluation tool, these measures are problematic
because their values are not explicitly related to segmentation errors.
In this talk we describe two approaches to rectify these problems. The first
approach provides a meaningful, quantitative assessment of segmentation
quality, in precision/recall terms, applicable until now only for supervised
evaluation. It characterizes the segments by a mixture of basic feature
distributions and uses a nonnegative matrix factorization (NMF) process to find
the components and the implied precision/recall estimates. A new NMF algorithm,
compatible with the earth moverís distance (EMD), is also developed. The
accuracy of the precision/recall estimates and the ability to optimize,
on-line, the segmentation algorithm parameters, are demonstrated.
The second approach proposes to characterize the segmentation by a rich set of
properties and uses a classifier to merge them into a scalar quality measure.
It uses several online and offline learning processes to optimize this measure,
which indeed characterizes segment quality better than the simple methods. A
search-based optimization leads to state-of-the-art segmentation.
This work was done together with Roman Sandler and David Peles.