Day54 - CBAM+NWD改进在FLIR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FLIR上训练200epochs
Yolov5s-cbam-nwd
Parameters: 9.502M
Class Precision Recall mAP50 mAP50-95 all 0.681 0.515 0.565 0.343 person 0.857 0.661 0.787 0.459 bike 0.593 0.512 0.507 0.314 car 0.879 0.775 0.859 0.62 motor 0.817 0.648 0.753 0.388 bus 0.783 0.531 0.678 0.477 truck 0.441 0.391 0.277 0.189 light 0.781 0.472 0.587 0.275 other vehicle 0.3 0.127 0.0739 0.0256 Yolov5m-cbam-nwd
Parameters: 25.896M
Class Precision Recall mAP50 mAP50-95 all 0.631 0.55 0.577 0.363 person 0.843 0.716 0.81 0.49 bike 0.573 0.571 0.55 0.361 car 0.857 0.813 0.878 0.647 motor 0.709 0.618 0.733 0.399 bus 0.748 0.587 0.679 0.493 truck 0.354 0.435 0.273 0.182 light 0.796 0.547 0.629 0.303 other vehicle 0.171 0.111 0.0652 0.0316 Yolov8s-cbam-nwd
Parameters: 11.516M
Class Precision Recall mAP50 mAP50-95 all 0.676 0.507 0.558 0.34 person 0.865 0.676 0.8 0.47 bike 0.651 0.524 0.538 0.338 car 0.89 0.779 0.867 0.63 motor 0.824 0.596 0.669 0.356 bus 0.807 0.542 0.683 0.471 truck 0.405 0.391 0.255 0.158 light 0.79 0.483 0.575 0.264 other vehicle 0.175 0.0635 0.0772 0.0359
本文由作者按照 CC BY 4.0 进行授权