Performance evaluation of probability density estimators for unsupervised information theoretical region merging

Abstract

Information theoretical region merging techniques have been shown to provide a state-of-the-art unified solution for natural and texture image segmentation. Here, we study how the segmentation results can be further improved by a more accurate estimation of the statistical model characterizing the regions. Concretely, we explore four density estimators that can be used for pdf or joint pdf estimation. The first three are based on different quantization strategies: a general uniform quantization, an MDL-based uniform quantization, and a data-dependent partitioning and estimation. The fourth strategy is based on a computationally efficient kernel-based estimator (averaged shifted histogram). Finally, all estimators are objectively evaluated using a database with available ground truth partitions.