Triplet Loss:
x
i
a
x_i^a
xia是anchor,
x
i
p
x_i^p
xip是positive ,
x
i
n
x_i^n
xin是negtive
∑
i
N
[
∣
∣
f
(
x
i
a
)
−
f
(
x
i
p
)
∣
∣
2
−
∣
∣
f
(
x
i
a
)
−
f
(
x
i
n
)
∣
∣
2
+
α
]
+
\sum_i^N[||f(x_i^a)-f(x_i^p)||^2 - || f(x_i^a)-f(x_i^n)||^2 + \alpha]_{+}
∑iN[∣∣f(xia)−f(xip)∣∣2−∣∣f(xia)−f(xin)∣∣2+α]+
Center Loss
在cross-entry loss的基础之上，增加同一类的均值loss.
L
=
L
S
+
λ
L
C
=
−
∑
i
=
1
m
l
o
g
e
W
y
i
T
x
i
+
b
y
i
∑
j
=
1
n
e
W
y
j
T
x
i
+
b
y
j
+
λ
2
∑
i
=
1
m
∣
∣
x
i
−
c
y
i
∣
∣
2
2
L=L_S+\lambda L_C = -\sum_{i=1}^m log \frac{e^{W_{y_i}^T x_i + b_{y_i}}}{\sum_{j=1}^n e^{W_{y_j}^T x_i + b_{y_j}}} + \frac{\lambda}{2}\sum_{i=1}^m ||x_i-c_{y_i}||^2_2
L=LS+λLC=−∑i=1mlog∑j=1neWyjTxi+byjeWyiTxi+byi+2λ∑i=1m∣∣xi−cyi∣∣22
A-softmax Loss
Large Angle Margin
L
a
n
g
=
1
N
∑
i
=
1
−
l
o
g
(
e
∣
∣
x
i
∣
∣
c
o
s
(
m
θ
y
i
,
i
)
e
∣
∣
x
i
∣
∣
c
o
s
(
m
θ
y
i
,
i
)
+
∑
j
≠
y
i
e
∣
∣
x
i
∣
∣
c
o
s
(
θ
y
i
,
i
)
)
L_{ang}=\frac{1}{N}\sum_{i=1}-log(\frac{e^{||x_i||cos(m \theta_{y_i},i)}}{e^{||x_i||cos(m \theta_{y_i},i)} + \sum_{j \neq y_i}e^{||x_i||cos(\theta_{y_i},i)}})
Lang=N1∑i=1−log(e∣∣xi∣∣cos(mθyi,i)+∑j=yie∣∣xi∣∣cos(θyi,i)e∣∣xi∣∣cos(mθyi,i))
扫一扫
12万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
