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

  Class	Precision	Recall	mAP50	mAP50-95
  all	0.502	0.385	0.388	0.227
  pedestrian	0.531	0.388	0.415	0.185
  people	0.559	0.266	0.316	0.118
  bicycle	0.271	0.158	0.122	0.0484
  car	0.719	0.765	0.787	0.554
  van	0.53	0.42	0.444	0.306
  truck	0.508	0.367	0.357	0.231
  tricycle	0.393	0.3	0.274	0.148
  awning-tricycle	0.292	0.201	0.146	0.0897
  bus	0.683	0.542	0.575	0.399
  motor	0.534	0.44	0.443	0.192

• Yolov5m-cbam-nwd

  Parameters: 25.896M

  Class	Precision	Recall	mAP50	mAP50-95
  all	0.527	0.412	0.422	0.253
  pedestrian	0.554	0.416	0.449	0.208
  people	0.574	0.285	0.338	0.129
  bicycle	0.285	0.179	0.151	0.0664
  car	0.746	0.779	0.805	0.578
  van	0.538	0.453	0.474	0.332
  truck	0.548	0.387	0.41	0.272
  tricycle	0.435	0.343	0.321	0.18
  awning-tricycle	0.325	0.205	0.169	0.0981
  bus	0.727	0.594	0.63	0.451
  motor	0.542	0.476	0.472	0.214

• Yolov8s-cbam-nwd

  Parameters: 11.516M

  Class	Precision	Recall	mAP50	mAP50-95
  all	0.506	0.386	0.395	0.232
  pedestrian	0.535	0.385	0.418	0.187
  people	0.547	0.267	0.32	0.118
  bicycle	0.266	0.173	0.132	0.0553
  car	0.735	0.765	0.792	0.56
  van	0.529	0.445	0.455	0.314
  truck	0.521	0.367	0.376	0.242
  tricycle	0.402	0.313	0.285	0.151
  awning-tricycle	0.315	0.188	0.151	0.0907
  bus	0.665	0.514	0.576	0.405
  motor	0.543	0.444	0.444	0.193