Methodology for extensive evaluation of semiautomatic and interactive segmentation algorithms using simulated Interaction models
Date
2016-09-21
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
ORCID
Type
Thesis
Degree Level
Doctoral
Abstract
Performance of semiautomatic and interactive segmentation(SIS) algorithms are usually evaluated by employing a small number of human operators to segment the images. The human operators typically provide the approximate location of objects of interest and their boundaries in an interactive phase, which is followed by an automatic phase where the segmentation is performed under the constraints of the operator-provided guidance. The segmentation results produced from this small set of interactions do not represent the true capability and potential of the algorithm being evaluated. For example, due to inter-operator variability, human operators may make choices that may provide either overestimated or underestimated results. As well, their choices may not be realistic when compared to how the algorithm is used in the field, since interaction may be influenced by operator fatigue and lapses in judgement. Other drawbacks to using human operators to assess SIS algorithms, include: human error, the lack of available expert users, and the expense. A methodology for evaluating segmentation performance is proposed here which uses simulated Interaction models to programmatically generate large numbers of interactions to ensure the presence of interactions throughout the object region. These interactions are used to segment the objects of interest and the resulting segmentations are then analysed using statistical methods. The large number of interactions generated by simulated interaction models capture the variabilities existing in the set of user interactions by considering each and every pixel inside the entire region of the object as a potential location for an interaction to be placed with equal probability. Due to the practical limitation imposed by the enormous amount of computation for the enormous number of possible interactions, uniform sampling of interactions at regular intervals is used to generate the subset of all possible interactions which still can represent the diverse pattern of the entire set of interactions.
Categorization of interactions into different groups, based on the position of the interaction inside the object region and texture properties of the image region where the interaction is located, provides the opportunity for fine-grained algorithm performance analysis based on these two criteria. Application of statistical hypothesis testing make the analysis more accurate, scientific and reliable in comparison to conventional evaluation of semiautomatic segmentation algorithms. The proposed methodology has been demonstrated by two case studies through implementation of seven different algorithms using three different types of interaction modes making a total of nine segmentation applications to assess the efficacy of the methodology. Application of this methodology has revealed in-depth, fine details about the performance of the segmentation algorithms which currently existing methods could not achieve due to the absence of a large, unbiased set of interactions. Practical application of the methodology for a number of algorithms and diverse interaction modes have shown its feasibility and generality for it to be established as an appropriate methodology. Development of this methodology to be used as a potential application for automatic evaluation of the performance of SIS algorithms looks very promising for users of image segmentation.
Description
Keywords
extensive evaluation, semiautomatic segmentation, inter-user variability, uniform sampling of interactions
Citation
Degree
Doctor of Philosophy (Ph.D.)
Department
Computer Science
Program
Computer Science