Chinese Journal of Liquid Crystals and Displays, Volume. 37, Issue 4, 539(2022)

YOLOv3 object detection method by introducing Gaussian mask self-attention module

Ya-jie KONG1,2 and Ye ZHANG1、*
Author Affiliations
  • 1State Key Laboratory of Applied Optics,Changchun Institute of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130033,China
  • 2University of Chinese Academy of Sciences,Beijing 100049,China
  • show less
    Figures & Tables(10)
    Architecture of YOLOv3-GMSA network
    Multi-scale feature fusion of feature pyramid network
    Structure of self-attention
    Structure of GMSA module
    Training curve of pivotal parameters in YOLOv3-GMSA
    Comparison of detection results
    • Table 1. Feature map size and prior frame size with picture size of 640×640

      View table
      View in Article

      Table 1. Feature map size and prior frame size with picture size of 640×640

      特征图尺寸感受野先验框尺寸
      20×20(116,90)
      (156,198)
      (373,326)
      40×40(30,61)
      (62,45)
      (59,119)
      80×80(10,13)
      (16,30)
      (33,23)
    • Table 2. Training environment

      View table
      View in Article

      Table 2. Training environment

      配置名称型号、参数
      CPUIntel(R)Core(TM)i9-9900K,8核
      固态硬盘金士顿,512 G
      内存金士顿,16 Gx2,8 Gx2
      显卡NVIDIA TITAN Xp,显存12 G,CUDA 10.2
      操作系统Ubuntu 18.04
      程序语言Python 3.8.11
      机器学习框架PyTorch 1.9.0
    • Table 3. Training parameters

      View table
      View in Article

      Table 3. Training parameters

      参数名称参数值
      批处理大小10
      迭代次数100
      学习率0.01
      动量0.937
      置信度阈值0.5
      NMS阈值0.5
      类别80
      自注意力头数量8
    • Table 4. Performance evaluation

      View table
      View in Article

      Table 4. Performance evaluation

      算法模型mAP@0.5/%mAP@0.5∶0.95/%Precision/%Recall/%FPS
      Faster R-CNN58.9838.5468.6756.3321.12
      SSD47.3329.4956.1944.9335.56
      ASSD52.7331.8160.2449.4836.71
      YOLOv354.3234.1761.7851.9643.26
      YOLOv3-GMSA56.8836.1965.3153.1839.38
    Tools

    Get Citation

    Copy Citation Text

    Ya-jie KONG, Ye ZHANG. YOLOv3 object detection method by introducing Gaussian mask self-attention module[J]. Chinese Journal of Liquid Crystals and Displays, 2022, 37(4): 539

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category:

    Received: Sep. 30, 2021

    Accepted: --

    Published Online: Jun. 20, 2022

    The Author Email: Ye ZHANG (yolanda@sp.irits.ai)

    DOI:10.37188/CJLCD.2021-0250

    Topics