一篇关于人脸识别中loss改进的方法
文章i地址：https://128.84.21.199/abs/1812.11317
作者github地址：https://github.com/xiaoboCASIA/SV-X-Softmax还未完成
有一个博主做了一个除了实验的全文翻译，想要看中文的可以去
calvinpaean：Support Vector Guided Softmax Loss for Face Recognition 论文学习
摘要
Face recognition has witnessed significant progresses due to the advances of deep convolutional neural networks (CNNs), the central challenge of which, is feature discrimination. To address it, one group tries to exploit mining-based strategies (e.g., hard example mining and focal loss) to focus on the informative examples. The other group devotes to designing margin-based loss functions (e.g., angular, additive and additive angular margins) to increase the feature margin from the perspective of ground truth class. Both of them have been well-verified to learn discriminative features. However, they suffer from either the ambiguity of hard examples or the lack of discriminative power of other classes. In this paper, we design a novel loss function, namely support vector guided softmax loss (SV-Softmax), which adaptively emphasizes the mis-classified points (support vectors) to guide the discriminative features learning. So the developed SV-Softmax loss is able to eliminate the ambiguity of hard examples as well as absorb the discriminative power of other classes, and thus results in more discrimiantive features. To the best of our knowledge, this is the first attempt to inherit the advantages of mining-based and margin-based losses into one framework. Experimental results on several benchmarks have demonstrated the effectiveness of our approach over state-of-the-arts.
作者分析了人脸识别中两种从loss方面促进feature discriminative的方法：mining-based和margin-based loss function，这两种方法都存在其局限性，作者提出了将两种方法结合的方法SV-Softmax loss。
main contribution
- We propose a novel SV-Softmax loss, which eliminates the ambiguity of hard examples as well as absorbs the discriminative power of other classes by focusing on support vectors. To the best of our knowledge, this is the first attempt to semantically fuse the mining-based and margin-based losses into one framework.
- We deeply analyze the relations of our SV-Softmax loss to the current mining-based and margin-based losses, and further develop an improved version SV-X-Softmax loss to enhance the feature discrimiantion.
作者仔细分析了提出的loss和focal loss、arc loss之间的区别和联系，这一部分分析的很好，是文章理论分析的亮点。 - We conduct extensive experiments on the benchmarks of LFW [8], MegaFace Challenge [9, 16] and Trillion Pairs Challenge, which have verified the superiority of our new approach over the baseline Softmax loss, the mining-based Softmax losses, the margin-based Soft-max losses, and their naive fusions.
作者做了大量的实验，实验的细节写的很多。
SV-Softmax loss
放几个公式和图来说明作者的思想
Softmax loss
L
1
=
−
l
o
g
e
s
c
o
s
(
θ
w
y
,
x
)
e
s
c
o
s
(
θ
w
y
,
x
)
+
∑
k
≠
y
K
e
s
c
o
s
(
θ
w
k
,
x
)
\mathcal L_1=-log\frac{e^{scos(\theta _{w_y,x})}}{e^{scos(\theta _{w_y,x})}+\sum_{k\neq y}^{K}e^{scos(\theta _{w_k,x})}}
L1=−logescos(θwy,x)+∑k̸=yKescos(θwk,x)escos(θwy,x)
Mining-based Softmax
L
2
=
−
g
(
p
y
)
l
o
g
e
s
c
o
s
(
θ
w
y
,
x
)
e
s
c
o
s
(
θ
w
y
,
x
)
+
∑
k
≠
y
K
e
s
c
o
s
(
θ
w
k
,
x
)
\mathcal L_2=-g(p_y)log\frac{e^{scos(\theta _{w_y,x})}}{e^{scos(\theta _{w_y,x})}+\sum_{k\neq y}^{K}e^{scos(\theta _{w_k,x})}}
L2=−g(py)logescos(θwy,x)+∑k̸=yKescos(θwk,x)escos(θwy,x)
mining-based softmax比softmax多了一个
g
(
p
y
)
g(p_y)
g(py)，利用这个
g
(
p
y
)
g(p_y)
g(py)来对hard samples着重进行训练。在Focal loss中，
g
(
p
y
)
=
(
1
−
p
y
)
γ
g(p_y)=(1-p_y)^\gamma
g(py)=(1−py)γ;在HM-Softmax中，hard sample时
g
(
p
y
)
=
1
g(p_y)=1
g(py)=1，否则
g
(
p
y
)
=
0
g(p_y)=0
g(py)=0。
Margin-based loss
根据arcface、sphereface、a-softmax几个loss的改进结果
L
3
=
−
l
o
g
e
s
f
(
m
,
θ
w
y
,
x
)
)
e
s
f
(
m
,
θ
w
y
,
x
)
+
∑
k
≠
y
K
e
s
c
o
s
(
θ
w
k
,
x
)
\mathcal L_3=-log\frac{e^{sf(m,\theta_{w_y,x}))}}{e^{sf(m,\theta_{w_y,x})}+\sum_{k\neq y}^{K}e^{scos(\theta _{w_k,x})}}
L3=−logesf(m,θwy,x)+∑k̸=yKescos(θwk,x)esf(m,θwy,x))
上式中，
f
(
m
,
θ
w
y
,
x
)
f(m,\theta_{w_y},x)
f(m,θwy,x)是精心设计的margin function.
Naive Mining-Margin Softmax loss
L
4
=
−
g
(
p
y
)
)
l
o
g
e
s
f
(
m
,
θ
w
y
,
x
)
)
e
s
f
(
m
,
θ
w
y
,
x
)
+
∑
k
≠
y
K
e
s
c
o
s
(
θ
w
k
,
x
)
\mathcal L_4=-g(p_y))log\frac{e^{sf(m,\theta_{w_y,x}))}}{e^{sf(m,\theta_{w_y,x})}+\sum_{k\neq y}^{K}e^{scos(\theta _{w_k,x})}}
L4=−g(py))logesf(m,θwy,x)+∑k̸=yKescos(θwk,x)esf(m,θwy,x))
Support Vector Guided Softmax Loss
L
5
=
−
g
(
p
y
)
)
l
o
g
e
s
c
o
s
(
θ
w
y
,
x
)
e
s
c
o
s
(
θ
w
y
,
x
)
+
∑
k
≠
y
K
h
(
t
,
θ
w
k
,
x
,
I
k
)
e
s
c
o
s
(
θ
w
k
,
x
)
\mathcal L_5=-g(p_y))log\frac{e^{scos(\theta_{w_y,x})}}{e^{scos(\theta_{w_y,x})}+\sum_{k\neq y}^{K}h(t,\theta_{w_k,x},I_k)e^{scos(\theta _{w_k,x})}}
L5=−g(py))logescos(θwy,x)+∑k̸=yKh(t,θwk,x,Ik)escos(θwk,x)escos(θwy,x)
其中，
h
(
t
,
θ
w
k
,
x
,
I
k
)
=
e
s
(
t
−
1
(
c
o
s
(
θ
w
k
,
x
)
+
1
)
I
k
h(t,\theta_{w_k},x,I_k)=e^{s(t-1(cos(\theta_{w_k,x})+1)I_k}
h(t,θwk,x,Ik)=es(t−1(cos(θwk,x)+1)Ik
当
t
=
1
t=1
t=1时，上式就等价于softmax
SV-X-Softmax
L
5
=
−
g
(
p
y
)
)
l
o
g
e
s
f
(
m
,
θ
w
y
,
x
)
e
s
f
(
m
,
θ
w
y
,
x
)
+
∑
k
≠
y
K
h
(
t
,
θ
w
k
,
x
,
I
k
)
e
s
c
o
s
(
θ
w
k
,
x
)
\mathcal L_5=-g(p_y))log\frac{e^{sf(m,\theta_{w_y,x})}}{e^{sf(m,\theta_{w_y,x})}+\sum_{k\neq y}^{K}h(t,\theta_{w_k,x},I_k)e^{scos(\theta _{w_k,x})}}
L5=−g(py))logesf(m,θwy,x)+∑k̸=yKh(t,θwk,x,Ik)escos(θwk,x)esf(m,θwy,x)
其中，
h
(
t
,
θ
w
k
,
x
,
I
k
)
=
e
s
(
t
−
1
(
c
o
s
(
θ
w
k
,
x
)
+
1
)
I
k
h(t,\theta_{w_k},x,I_k)=e^{s(t-1(cos(\theta_{w_k,x})+1)I_k}
h(t,θwk,x,Ik)=es(t−1(cos(θwk,x)+1)Ik
扫一扫
1184
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
