conda create -n fcos python==3.7.
pip install torch torchvision.
pip install opencv-python.
pip install TensorBoard.
use VOC format, and data can download in https://pan.baidu.com/s/1hssdO_I7vFnSyIw6GuzUBw Password: 6yr1.
can modify in model/config.py.
the backbone support choice in resnet18 and vovnet39.
python train.py.
Focal loss for class:
people who wear mask will be mark face_mask label.
people who not wear mask will be mask face label only.The project was completed by me independently for academic exchange. For commercial use, please contact me by email an_chao1994@163.com.
the code refer ro:https://github.com/VectXmy/FCOS.Pytorch.
I find use GN in head is better.
I find use deformable_conv in head is better, I only instead of the first general_conv in head. refer ro:https://github.com/ChunhuanLin/deform_conv_pytorch.
I find use ELU in head instead of Relu is better.
- Please show love to me who is live in the poor mountainous area(optional)