文章

Day53 - CBAM+NWD改进在VisDrone上的表现

yolov5s/m和yolov8s在backbone的C3/C2f模块后添加一层CBAM,使用NWD损失函数,在VisDrone2019上训练200epochs

  • Yolov5s-cbam-nwd

    Parameters: 9.502M

    ClassPrecisionRecallmAP50mAP50-95
    all0.5020.3850.3880.227
    pedestrian0.5310.3880.4150.185
    people0.5590.2660.3160.118
    bicycle0.2710.1580.1220.0484
    car0.7190.7650.7870.554
    van0.530.420.4440.306
    truck0.5080.3670.3570.231
    tricycle0.3930.30.2740.148
    awning-tricycle0.2920.2010.1460.0897
    bus0.6830.5420.5750.399
    motor0.5340.440.4430.192
  • Yolov5m-cbam-nwd

    Parameters: 25.896M

    ClassPrecisionRecallmAP50mAP50-95
    all0.5270.4120.4220.253
    pedestrian0.5540.4160.4490.208
    people0.5740.2850.3380.129
    bicycle0.2850.1790.1510.0664
    car0.7460.7790.8050.578
    van0.5380.4530.4740.332
    truck0.5480.3870.410.272
    tricycle0.4350.3430.3210.18
    awning-tricycle0.3250.2050.1690.0981
    bus0.7270.5940.630.451
    motor0.5420.4760.4720.214
  • Yolov8s-cbam-nwd

    Parameters: 11.516M

    ClassPrecisionRecallmAP50mAP50-95
    all0.5060.3860.3950.232
    pedestrian0.5350.3850.4180.187
    people0.5470.2670.320.118
    bicycle0.2660.1730.1320.0553
    car0.7350.7650.7920.56
    van0.5290.4450.4550.314
    truck0.5210.3670.3760.242
    tricycle0.4020.3130.2850.151
    awning-tricycle0.3150.1880.1510.0907
    bus0.6650.5140.5760.405
    motor0.5430.4440.4440.193
本文由作者按照 CC BY 4.0 进行授权