FRAMEWORK FOR COMPARING SEGMENTATION ALGORITHMS

Abstract. The notion of a ‘Best’ segmentation does not exist. A segmentation algorithm is chosen based on the features it yields, the properties of the segments (point sets) it generates, and the complexity of its algorithm. The segmentation is then assessed based on a variety of metrics such as homogeneity, heterogeneity, fragmentation, etc. Even after an algorithm is chosen its performance is still uncertain because the landscape/scenarios represented in a point cloud have a strong influence on the eventual segmentation. Thus selecting an appropriate segmentation algorithm is a process of trial and error.

Automating the selection of segmentation algorithms and their parameters first requires methods to evaluate segmentations. Three common approaches for evaluating segmentation algorithms are ‘goodness methods’, ‘discrepancy methods’ and ‘benchmarks’. Benchmarks are considered the most comprehensive method of evaluation. This paper shortcomings in current benchmark methods are identified and a framework is proposed that permits both a visual and numerical evaluation of segmentations for different algorithms, algorithm parameters and evaluation metrics. The concept of the framework is demonstrated on a real point cloud. Current results are promising and suggest that it can be used to predict the performance of segmentation algorithms.