Evaluating Segmentation Quality via Reference Segmentations in Tree-like Structure

Abstract

In the image segmentation task, different understandings of the image content will lead to different granularities of segmentation results. Existing segmentation evaluation methods generally use one or more reference segmentations to evaluate the quality of image segmentation. But the limited number of reference segmentations can not give an comprehensive definition on the image granularity division. To solve the this problem, we present a segmentation evaluation method based on tree structure. Firstly, the regional granularity analysis is performed on multiple reference segmentations of the same image. A multilevel region tree is constructed and different layers in the region tree will correspond to different granularities of the reference segmentations; Secondly, for a segmentation to be evaluated, we adaptively select a layer in the region tree as a reference segmentation, which has similar region granularity with the input segmentation. The proposed evaluation method utilizes multilevel information in the image content, which leads to a more accurate and objective evaluation.