Title: Optimizing the parameters of image segmentation with evolution strategies 

Abstract
Despite the volume and variety of image segmentation techniques available,
relatively little quantitative evaluation of them has been done. Rather,
authors usually present their results visually and argue that the results
demonstrate an improvement over existing methods. One exception is a
Berkeley study comprised of a database of human-segmented images and an
algorithm for comparing the results of automatic segmentations with the
human 'benchmarks'. However, not all applications of image segmentation
require agreement with human intuition. For example in object-based video
coding the most important requirement is that segmentations of successive
frames are spatially stable; the human benchmark is an indirect evaluator
in this case. We previously presented an index for evaluating the
stability of a segmentation algorithm. This seminar presents the
results of an experiment using this index to optimize the merging
criterion of a region-merging segmentation algorithm. Optimization is
performed with respect to stability under noise and under slight changes
in camera position. The importance of these segmentation qualities in
computer vision applications and the Evolution Strategies method for
gradient ascent will be discussed.