文章

Day55 - P234+CBAM改进在VisDrone2019上的表现

yolov5s/m和yolov8s,删除P5检测头+添加P2检测头,在backbone的C3/C2f模块后添加一层CBAM,在VisDrone2019上训练200epochs

  • Yolov5s-p234-cbam

    Parameters: 8.626M

    ClassPrecisionRecallmAP50mAP50-95
    all0.5260.4230.4360.266
    pedestrian0.5720.480.5140.248
    people0.570.3760.4140.17
    bicycle0.2980.190.1640.0778
    car0.7210.8230.8390.603
    van0.5360.4780.4880.347
    truck0.5070.3510.3680.252
    tricycle0.450.2980.3070.176
    awning-tricycle0.3270.1940.1630.108
    bus0.7050.5420.5890.441
    motor0.570.4990.510.239
  • Yolov5m-p234-cbam

    Parameters: 23.546M

    ClassPrecisionRecallmAP50mAP50-95
    all0.5620.4490.4670.286
    pedestrian0.6150.4930.5420.266
    people0.5930.3940.4380.185
    bicycle0.3360.2140.190.0897
    car0.7640.8240.850.619
    van0.5660.4830.5050.368
    truck0.6030.4070.4260.293
    tricycle0.4980.350.3480.201
    awning-tricycle0.3380.2110.1920.121
    bus0.7070.5820.6280.457
    motor0.5950.5310.5470.264
  • Yolov8s-p234-cbam

    Parameters: 10.564M

    ClassPrecisionRecallmAP50mAP50-95
    all0.5430.4290.4460.272
    pedestrian0.5970.4740.5210.252
    people0.5770.3840.4260.179
    bicycle0.3050.2040.1690.0792
    car0.7470.8190.8420.611
    van0.5730.4740.4930.356
    truck0.5480.3630.3780.256
    tricycle0.4390.3150.3050.177
    awning-tricycle0.3730.2070.1840.121
    bus0.6820.550.6170.444
    motor0.5870.5020.5240.248
本文由作者按照 CC BY 4.0 进行授权