Real-time maximally stable homogeneous regions

Abstract

In this paper, we present the concept of maximally stable homogeneous regions (MSHR). MSHR are conceptually very similar to maximally stable extremal regions but can be segmented in images with an arbitrary number of channels. The computation of the presented MSHR relies on the construction of a quasi-flat zone hierarchy. We present a fast algorithm for computing the hierarchy that overcomes the runtime restrictions of existing approaches. The proposed algorithm can construct the quasi-flat zone hierarchy efficiently in real time, scales linearly in the number of pixels and, in practice, sub-linearly in the number of channels. In the experiments, we display how MSHR can be used to improve the results of optical character recognition systems and to perform 3D object segmentation. We further demonstrate the universality and speed of the proposed algorithm for three example applications: image segmentation, object tracking, and image filtering.