Statistical modeling for minimum -effort interactive boundary extraction

Christina Pavlopoulou, Purdue University

Abstract

Segmentation is an important computer vision problem, however for most realistic situations it cannot be solved reliably. Interactive segmentation is a desirable alternative especially for human-in-the-loop applications. Although several interactive segmentation methods have been proposed in the past and the problem is gaining increasing popularity there are several issues that have not received sufficient attention if at all. This thesis focuses on what we believe is an important issue for an interactive segmentation method, namely, how does one define an interactive segmentation criterion capable of extracting the boundary of an object with small amount of human input. We answer this question for the case of criteria that can be optimized with shortest paths algorithms and in particular Dijkstra’s algorithm. Such criteria are discontinuity seeking and the type of human input they accept consists of points on the boundary of the object to be segmented. Existing criteria that can be optimized with shortest paths have a bias towards contours of small length, thus increasing the amount of human input required to delineate a structure. We propose a novel statistical criterion that aims to address this issue. The key quantity that differentiates our criterion from existing work is a normalization factor that boosts the probability of locally best segments. This property enables the efficient, both computationally and in terms of human input, extraction of geometrically complex boundaries, something that was not possible before for this type of optimization methods.

Degree

Ph.D.

Advisors

Kak, Purdue University.

Subject Area

Computer science

COinS