Interactive Object Segmentation using Binary Inputs

Abstract

Every day, humans use their vision to process millions of pixels and select regions of interest. This task of highlighting and grouping pixels of interest in a scene is called image segmentation, and it is a fundamental method that humans use to communicate with each other ideas, concepts, and emotions. We introduce a method derived from feedback information theory that allows individuals with motor control disabilities to perform image segmentation using only binary inputs and a simple encoding scheme. We propose two versions of our algorithm, and evaluate their ability to specify desired regions for the user with restricted inputs and noise on large, publicly available image data sets. We also compare our method to the previous best algorithm, developed by Rupprecht et al.