文章

Day34 - NWD在FLIR上的表现

yolov5n/s/m和yolov8n/s仅使用NWD损失函数,在FLIR上训练200epochs

  • Yolov5n-nwd

    Parameters: 2.655M

    ClassPrecisionRecallmAP50mAP50-95
    all0.5990.4640.4910.286
    person0.8040.6290.7350.402
    bike0.4670.4470.4230.252
    car0.8320.7480.8250.574
    motor0.7990.5780.6580.333
    bus0.8170.450.6070.406
    truck0.2730.4130.1640.0902
    light0.720.4180.4910.219
    other vehicle0.07710.03170.0250.00881
  • Yolov5s-nwd

    Parameters: 9.153M

    ClassPrecisionRecallmAP50mAP50-95
    all0.6510.5070.5550.338
    person0.860.6510.7810.453
    bike0.5560.5180.5020.296
    car0.8760.7780.8630.622
    motor0.7450.5820.7010.384
    bus0.7940.5360.6480.447
    truck0.3780.4350.2940.193
    light0.8020.4580.5660.268
    other vehicle0.1980.09520.08240.0423
  • Yolov5m-nwd

    Parameters: 25.111M

    ClassPrecisionRecallmAP50mAP50-95
    all0.6810.5170.5790.364
    person0.8840.6570.8010.486
    bike0.6260.5410.5540.356
    car0.9060.7860.8740.642
    motor0.7970.6360.7360.386
    bus0.8140.5030.690.501
    truck0.3770.4570.30.213
    light0.8520.4940.6230.298
    other vehicle0.190.06350.05830.0295
  • Yolov8n-nwd

    Parameters: 3.157M

    ClassPrecisionRecallmAP50mAP50-95
    all0.6120.4280.490.292
    person0.8480.6130.7420.411
    bike0.5630.4470.440.262
    car0.8640.7380.8330.584
    motor0.7580.5450.6430.33
    bus0.7520.4580.6120.431
    truck0.2260.2160.1350.0924
    light0.8020.3740.4950.219
    other vehicle0.08430.03170.02410.0103
  • Yolov8s-nwd

    Parameters: 11.167M

    ClassPrecisionRecallmAP50mAP50-95
    all0.6330.5010.5430.335
    person0.8490.6690.7840.457
    bike0.5670.5410.5140.312
    car0.8780.7790.8630.621
    motor0.7480.60.6710.373
    bus0.7710.5590.660.47
    truck0.2690.3040.2080.134
    light0.80.4740.580.27
    other vehicle0.1810.07940.06730.0416
本文由作者按照 CC BY 4.0 进行授权