pytorch-----RetinaFace(代码网络理解)

1、网络模型
pytorch-----RetinaFace(代码网络理解)_第1张图片
2、训练
代码在之前的博客中。
(1)骨干网:mobilenet0.25
输入图片[640,640,3]
batchsize=32
有三层输出:
对应图片的c3,c4,c5;
out1[32,64,80,80]
out2[32,128,40,40]
out3[32,256,20,20]
特征金字塔:
fpn1[32,64,80,80]
fpn2[32,64,40,40]
fpn3[32,64,20,20]
上下文模块
ssh1[32,64,80,80]
ssh2[32,64,40,40]
ssh3[32,64,20,20]
Multi-task loss:
bbox[32,16800,4]
class[32,16800,2]
landm[32,16800,10]
(2)骨干网:Resnet50
输入图片[640,640,3]
batchsize=4
有三层输出:
对应图片的c3,c4,c5;
out1[4,512,80,80]
out2[4,1024,40,40]
out3[4,2048,20,20]
特征金字塔:
fpn1[4,256,80,80]
fpn2[4,256,40,40]
fpn3[4,256,20,20]
上下文模块
ssh1[4,256,80,80]
ssh2[4,256,40,40]
ssh3[4,256,20,20]
Multi-task loss:
bbox[4,16800,4]
class[4,16800,2]
landm[4,16800,10]
3、测试
(1)骨干网:mobilenet0.25
输入图片[624,1024,3]

有三层输出:
对应图片的c3,c4,c5;
out1[1,64,78,128]
out2[1,128,39,64]
out3[1,256,20,32]
特征金字塔:
fpn1[1,64,78,128]
fpn2[1,64,39,64]
fpn3[1,64,20,32]
上下文模块
ssh1[1,64,78,128]
ssh2[1,64,39,64]
ssh3[1,64,20,32]
Multi-task loss:
bbox[1,26240,4]
class[1,26240,2]
landm[1,26240,10]
(2)骨干网:Resnet50
输入图片[624,1024,3]

有三层输出:
对应图片的c3,c4,c5;
out1[1,512,78,128]
out2[1,1024,39,64]
out3[1,2048,20,32]
特征金字塔:
fpn1[1,256,78,128]
fpn2[1,256,39,64]
fpn3[1,256,20,32]
上下文模块
ssh1[1,2256,78,128]
ssh2[1,256,39,64]
ssh3[1,256,20,32]
Multi-task loss:
bbox[1,26240,4]
class[1,26240,2]
landm[1,26240,10]

你可能感兴趣的:(RetinaFace)