Image Segmentation Stability: An Empirical Investigation
C. Scott Brown, William Bradley Glisson, Ryan Benton, Jordan Shropshire, Thomas Watts, Timothy Sullivan · 2018
This paper is an evaluation of the stability of image segmentation algorithms. We provide a definition of stability for segmentation algorithms, as well as a procedure for formulating stability measures in a given context. Using this procedure, a systematic empirical analysis of the stability of three popular image segmentation algorithms is performed. The stability of the algorithms is examined in the context of three real world transformations: video, compression and subimage alterations. The analysis provides evidence that stability is a meaningful property of an algorithm, and could be helpful for algorithm selection. This result lays the groundwork for developing methods for using stability as a means of automatic model and hyperparameter selection.