Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (6): 57-69.DOI: 10.3778/j.issn.1002-8331.2207-0139

• Research Hotspots and Reviews • Previous Articles     Next Articles

Survey of Evaluation Metrics and Methods for Semantic Segmentation

YU Ying, WANG Chunping, FU Qiang, KOU Renke, WU Weiyi, LIU Tianyong   

  1. 1.Department of Electronic and Optical Engineering, Shijiazhuang Campus, Army Engineering University of PLA, Shijiazhuang 050005, China
    2.School of Information and Intelligent Engineering, University of Sanya, Sanya, Hainan 572022, China
    3.Department of Equipment Command and Management, Shijiazhuang Campus, Army Engineering University of PLA, Shijiazhuang 050005, China
    4.School of Earth Sciences, Northeast Petroleum University, Daqing, Heilongjiang 163319, China
  • Online:2023-03-15 Published:2023-03-15

语义分割评价指标和评价方法综述

于营,王春平,付强,寇人可,吴巍屹,刘天勇   

  1. 1.陆军工程大学石家庄校区 电子与光学工程系,石家庄 050005
    2.三亚学院 信息与智能工程学院,海南 三亚 572022
    3.陆军工程大学石家庄校区 装备指挥与管理系,石家庄 050005
    4.东北石油大学 地球科学学院,黑龙江 大庆 163319

Abstract: Deep learning has made major breakthroughs in the field of semantic segmentation. Standard, well-known and comprehensive metrics should be used to evaluate the performance of these algorithms to ensure objectivity and effectiveness of the evaluation. Through summary of the existing semantic segmentation evaluation metrics, this paper elaborates from some aspects, e.g., pixel accuracy, depth estimation error metric, operation efficiency, memory demand and robustness. Especially, the widely used accuracy metrics such as F1 score, mIoU, mPA, Dice coefficient and Hausdorff distance are introduced in detail. In addition, this paper expounds the related research on the robustness and generalization. Furthermore, this paper points out the requirements in the semantic segmentation experiment and the limitations of segmentation quality evaluation.

Key words: semantic segmentation, evaluation metric, mean intersection over union(mIoU), mean pixel accuracy(mPA), robustness

摘要: 深度学习算法在语义分割领域已经取得大量突破,对这些算法的性能评估应选择标准、通用、全面的度量指标,以保证评价的客观性和有效性。通过对当前语义分割评价指标和度量方法进行归纳分析,从像素标记准确性、深度估计误差度量、执行效率、内存占用、鲁棒性等方面进行了多角度阐述,尤其对广泛应用的F1分数、mIoU、mPA、Dice系数、Hausdorff距离等准确性指标进行了详细介绍,并总结了提高分割网络鲁棒性的方法,指出了语义分割实验的要求和当前分割质量评价存在的问题。

关键词: 语义分割, 评价指标, 平均交并比(mIoU), 平均像素精度(mPA), 鲁棒性