Day51 - P234+NWD改进在VisDrone上的表现

yolov5s/m和yolov8s删除P5检测头+添加P2检测头,使用NWD损失函数,在VisDrone上训练200epochs

• Yolov5s-p234-nwd

  Parameters: 6.273M

  Class	Precision	Recall	mAP50	mAP50-95
  all	0.538	0.414	0.434	0.257
  pedestrian	0.584	0.472	0.511	0.24
  people	0.588	0.366	0.412	0.167
  bicycle	0.308	0.193	0.17	0.0737
  car	0.728	0.816	0.834	0.592
  van	0.551	0.462	0.472	0.334
  truck	0.502	0.336	0.362	0.236
  tricycle	0.478	0.316	0.314	0.173
  awning-tricycle	0.335	0.18	0.165	0.105
  bus	0.732	0.499	0.583	0.413
  motor	0.57	0.5	0.514	0.236

• Yolov5m-p234-nwd

  Parameters: 17.162M

  Class	Precision	Recall	mAP50	mAP50-95
  all	0.564	0.445	0.466	0.282
  pedestrian	0.633	0.493	0.551	0.265
  people	0.6	0.399	0.436	0.18
  bicycle	0.326	0.23	0.204	0.0933
  car	0.768	0.821	0.843	0.61
  van	0.569	0.476	0.495	0.353
  truck	0.59	0.38	0.412	0.272
  tricycle	0.483	0.355	0.341	0.192
  awning-tricycle	0.351	0.201	0.18	0.111
  bus	0.716	0.57	0.649	0.482
  motor	0.603	0.527	0.544	0.259

• Yolov8s-p234-nwd

  Parameters: 7.781M

  Class	Precision	Recall	mAP50	mAP50-95
  all	0.525	0.425	0.438	0.261
  pedestrian	0.58	0.478	0.519	0.246
  people	0.569	0.373	0.413	0.169
  bicycle	0.296	0.207	0.178	0.0786
  car	0.738	0.819	0.837	0.598
  van	0.548	0.466	0.48	0.34
  truck	0.519	0.343	0.361	0.23
  tricycle	0.423	0.303	0.298	0.166
  awning-tricycle	0.326	0.186	0.168	0.109
  bus	0.691	0.543	0.597	0.431
  motor	0.559	0.529	0.525	0.242