This repo implements PyramidBox with pytorch The paper PyramidBox is based on VGG16. But this repo is based on resnet50.
Here, I implements data-anchor-sampling, LFPN, max-in-out layer.
The result on WIDER FACE Val set:
AP Easy | AP Medium | AP Hard |
---|---|---|
95.3 | 94.3 | 89.0 |
You can download model from Baidu!
With the model and test.py, you can get the same result on WIDER FACE Val set!
*Python3
*Pytorch3.1
*OpenCV3
annoPath is path to your label file.
The label file should in the following format:
path_to_img1 num_face1 X1 Y1 W1 H1 X2 Y2 ... Wn1 Hn1
path_to_img2 num_face2 X1 Y1 W1 H1 X2 Y2 ... Wn2 Hn2