ps 640
YOLOV5 Param 1.9M~86.7M AP 28.0%~50.7% FPS 83~159
YOLOX Param 9.0M~99.1M AP 40.5%~51.5% FPS 58~102
YOLOR Param 52.9M~96.9M AP 50.8%~53.0% FPS 87~106
PPYOLOE Param 7.9M~98.4M AP 42.7%~52.2% FPS 45~208
YOLOV7 Param 6.2M~71.3M AP 38.7%~53.1% FPS 114~286
ps 1280
YOLOV5 Param 3.2M~140.7M AP 36.0%~55.0% FPS 38~123
YOLOR Param 37.2M~151.7M AP 53.5%~56.5% FPS 34~76
YOLOV7 Param 70.4M~154.7M AP 54.6%~56.8% FPS 36~84
YOLOV7
Backbone:
位置: cfg/training/yolov7.yaml
models/yolo.py
1.使用ELAN(包括复杂 E-ELAN结构)和MP结构
CSB=Conv->BN->silu激活
MP=(MaxPool->CSB) cat (CBS->CBS)->
ELAN=N个CBS->cat->CBS
ELAN结构:::
通过控制最短最长的梯度路径,更深的网络可以有效地学习和收敛。作者提出ELAN结构。基于ELAN设计的E-ELAN 用expand、shuffle、merge card