1. Prior Box layer
- m*n个cell
- 每个cell上生成固定scale和aspect ratio的box
1) 假设一个feature map有m*n个cell, 每个cell对应k个default box, 每个default box预测c个类别的score和4个offset(坐标)
- (c+4) km*n个输出
2. 参数定义
- Scale:
S
k
=
S
m
i
n
+
S
m
a
x
?
S
m
i
n
m
?
1
?
(
k
?
1
)
,
k
∈
[
1
,
m
]
S_k = S_{min}+ \frac{S_{max} - S_{min}}{m - 1}*(k-1), k \isin[1,m]
Sk?=Smin?+m?1Smax??Smin???(k?1),k∈[1,m]
- Smin = 0.2, 最底层scale是0.2; Smax =0.9,最高层scale是0、9
- Aspect ratio: ar = { 1, 2, 3, 1/2, 1/3}
- 宽:
w
k
a
=
s
k
?
a
r
w_k^a = s_k*\sqrt a_r
wka?=sk??a
?r?
- 高:
h
k
a
=
s
k
?
a
r
h_k^a = s_k*\sqrt a_r
hka?=sk??a
?r?
- Aspect ratio = 1, 增加一种scale的default box:
w
k
′
=
s
k
?
s
k
+
1
w'_k = \sqrt {s_k*s_{k+1}}
wk′?=sk??sk+1?
?
- 每个feature map cell 定义6种default box
|