New evaluation method for 3D mesh segmentation

摘要

Segmentation of 3D models plays a major role in many computer vision applications. Over the last few decades, extensive research has been done to propose many different segmentation algorithms. However, comparing the segmentation quality of multiple algorithms still a difficult task. Recently, many works have been proposed to solve this problem by comparing the segmentation resulting from an automatic algorithm, with a reference segmentation generally called a ground truth segmentation. As a result of this comparison a score of similarity/dissimilarity is extracted which represent the quality of the segmentation. Nevertheless, almost every 3D model has many reference segmentations, and comparing with just one ground truth cannot give a real representative score of the segmentation. In this paper we propose a new evaluation method inspired by the Jaro distance, which we adapt to evaluate 3D segmentation algorithms and to compare an automatic segmentation with a set of reference segmentations. Experimental results are reported in the last section of this paper to validate the proposed measure and compare it with others well-known measures.