Day52 - P234+NWD改进在FLIR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FLIR上训练200epochs
Yolov5s-p234-nwd
Parameters: 6.273M
Class Precision Recall mAP50 mAP50-95 all 0.606 0.538 0.564 0.346 person 0.838 0.736 0.837 0.501 bike 0.536 0.529 0.517 0.324 car 0.846 0.826 0.888 0.645 motor 0.681 0.636 0.686 0.356 bus 0.8 0.493 0.651 0.466 truck 0.251 0.457 0.188 0.118 light 0.784 0.583 0.669 0.331 other vehicle 0.109 0.0411 0.0786 0.0298 Yolov5m-p234-nwd
Parameters: 17.162M
Class Precision Recall mAP50 mAP50-95 all 0.662 0.557 0.597 0.38 person 0.88 0.729 0.856 0.531 bike 0.635 0.571 0.55 0.365 car 0.884 0.824 0.9 0.666 motor 0.708 0.691 0.717 0.393 bus 0.81 0.508 0.698 0.504 truck 0.325 0.413 0.247 0.163 light 0.797 0.592 0.682 0.345 other vehicle 0.258 0.127 0.128 0.0741 Yolov8s-p234-nwd
Parameters: 7.781M
Class Precision Recall mAP50 mAP50-95 all 0.663 0.496 0.566 0.349 person 0.894 0.667 0.839 0.504 bike 0.678 0.52 0.521 0.32 car 0.907 0.787 0.893 0.651 motor 0.692 0.571 0.631 0.333 bus 0.8 0.508 0.689 0.485 truck 0.327 0.348 0.239 0.155 light 0.836 0.522 0.654 0.32 other vehicle 0.17 0.0476 0.0579 0.0243
本文由作者按照 CC BY 4.0 进行授权