S3FD
6个尺度,核心网络vgg 16
160 80 40 20 10
anchors一共34125个
1070 batch-size只能4,速度30ms
人脸检测对比:
https://github.com/cs-giung/face-detection-pytorch
算法来自下面的:有权重,手检测和人脸检测,权重80多m
https://github.com/yxlijun/S3FD.pytorch
上面两个代码和下面这个网络好像是一样的,训练代码待看
无权重:有训练代码
https://github.com/luuuyi/S3FD.PyTorch
Subset | Original Caffe | PyTorch Implementation |
---|---|---|
Easy | 93.7% | 94.1% |
Medium | 92.4% | 92.9% |
Hard | 85.2% | 85.4% |
https://github.com/tgisaturday/S3FD_ATSS_SAPD
1070 batch-size只能4,速度30ms
net = build_s3fd('train', num_classes=2).cuda()
inputs = torch.randn(4, 3, 640, 640).cuda()
for i in range(10):
start=time.time()
output = net(inputs)
print('time',time.time()-start)
❝ 对max-out background label一个直观的解释就是对于每一个小尺寸anchor,进行N次人脸和背景分类,选择其中某个背景Acore最高的一个作为该Anchor的score。其实就是对Anchor进行多次预测,然后取其最大背景概率结果,以此降低误检为人脸的anchor数量,从而降低假阳性率。
https://github.com/jianzhnie/S3FD_pytorch
WIDER FACE:
Easy AP=0.925 paper = 0.927
Medium AP=0.925 paper = 0.924
Hard AP=0.854 paper = 0.852