中文

基于自适配归一化的改进Mask Scoring R-CNN

Hits:

  • Key Words:自适配归一化;批量归一化;目标检测;实例分割

  • Abstract:基于批量归一化的mask scoring R-CNN在目标检测与实例分割领域展现出卓越性能,其平均精度明显高于传统实例分割模型Mask R-CNN。但是由于批量归一化方法存在小批量精度骤降和大批量GPU内存溢出的缺陷,影响到实际应用中的检测与分割任务效果。自适配归一化方法对各批量大小都有极佳的鲁棒性,可以弥补上述不足。从数学角度给出了减少自适配归一化中计算冗余的证明,并将其应用于mask scoring R-CNN,小批量条件下在COCO数据集内将检测精度提升了4.4%,分割精度提升了3.9%,进一步提升了模型性能。

  • Volume:v.43;No.338

  • Issue:06

  • Translation or Not:no


Laoshan Campus-99 Songling Road, Qingdao City, Shandong Province
Sifang Campus-No.53 Zhengzhou Road, Qingdao City, Shandong Province
Sino-German International Cooperation Zone (Sino-German Campus)-No. 3698 Tuanjie Road, West Coast New District, Qingdao City, Shandong Province
Gaomi Campus-No. 1 Xingtan West Street, Gaomi City, Shandong Province
Jinan Campus-No. 80 Wenhua East Road, Jinan City, Shandong Province ©2015 Qingdao University of Science and Technology
Administrator email: master@qust.e
Click:
  MOBILE Version

The Last Update Time:..