文章

Day56 - P234+CBAM改进在FLIR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FLIR上训练200epochs

  • Yolov5s-p234-cbam

    Parameters: 8.626M

    ClassPrecisionRecallmAP50mAP50-95
    all0.6160.5380.5710.36
    person0.8560.7250.840.511
    bike0.5730.5290.510.326
    car0.8640.8190.8910.657
    motor0.6460.6320.7020.386
    bus0.7380.5040.6570.464
    truck0.2920.4350.2330.164
    light0.7870.5640.6570.329
    other vehicle0.170.09520.07880.042
  • Yolov5m-p234-cbam

    Parameters: 23.546M

  • Yolov8s-p234-cbam

    Parameters: 10.564M

    ClassPrecisionRecallmAP50mAP50-95
    all0.6560.5250.5770.363
    person0.8560.7210.8390.511
    bike0.6390.5530.5520.362
    car0.8760.8130.8930.659
    motor0.7080.5820.6440.347
    bus0.8370.5080.6950.488
    truck0.3770.4130.2660.19
    light0.7820.5610.6680.328
    other vehicle0.1720.04620.0560.0174
本文由作者按照 CC BY 4.0 进行授权