AAAI-2020
文章目录
1 Background and Motivation
为提升 CNN 模型的泛化能力，提升 CNN 模型应对遮挡场景的鉴别能力，提出了一种数据增广方法，Random Erasing——random position with random-sized mask with random pixel
While we can
manually
add occluded natural images to the training set, it is
costly
and the levels of occlusion might be
limited
.
2 Related Work
Dropout / DropConect / Adaptive dropout / Stochastic Pooling / DisturbLabel / PatchShuffle
Random flipping / random cropping
3 Advantages / Contributions
提出 Random Erasing Data Augmentation，与 random cropping，random filpping 互补，not require any extra parameter learning，在分类 / 检测 / reID 任务上均有不错的效果
4 Method
1）算法流程
x
e
x_e
x
e
和
y
e
y_e
y
e
是中心点，
W
e
W_e
W
e
和
H
e
H_e
H
e
是长宽
随机中心点，随机长宽比，随机面积，随机像素填充值
2）分类和 ReID 上的应用
简单粗暴，全图范围 random
3）目标检测上的应用
全图范围，目标范围，全图 + 目标范围
4）和 Random cropping 的区别
random cropping，可以降低背景的干扰，can base learning models on the presence of parts of the object instead of focusing on the whole object
random erasing，can be viewed as adding noise to the image
结合起来样本更加丰富
5 Experiments
5.1 Datasets and Metrics
1）数据集
-
分类
CIFAR-10
CIFAR-100
Fashion-MNIST -
检测
PASCAL VOC 2007 -
ReID
Market-1501
DukeMTMC-reID
CUHK03
2）评价指标
分类，top-1 error rates，“mean std” based on 5 runs
检测，mAP
ReID，rank-1，mAP
5.2 Experiments
5.2.1 Image Classification
1）Classification accuracy on different datasets
p
=
0.5
p = 0.5
p
=
0
.
5
,
s
l
=
0.02
s_l = 0.02
s
l
=
0
.
0
2
,
s
h
=
0.4
s_h = 0.4
s
h
=
0
.
4
, and
r
1
=
1
/
r
2
=
0.3
r1 =1/r2= 0.3
r
1
=
1
/
r
2
=
0
.
3
2）The impact of hyper-parameters
fix
s
l
s_l
s
l
to 0.02,
r
1
=
1
/
r
2
r1 = 1/r2
r
1
=
1
/
r
2
and evaluate
p
p
p
,
s
h
s_h
s
h
, and
r
1
r1
r
1
We set
p
=
0.5
p = 0.5
p
=
0
.
5
,
s
h
=
0.4
s_h = 0.4
s
h
=
0
.
4
and
r
1
=
0.3
r1 = 0.3
r
1
=
0
.
3
as the base setting. When evaluating one of the parameters, we fixed the other two parameters
都比 Baseline（没有 random erasing） 效果要更好！
3）Four types of random values for erasing
随机填充值和填充 ImageNet 的 mean 效果差不多[125, 122,114]（算是明白 114 的出处了，原来那个灰白颜色是 ImageNet 的均值），优于填 0 和 255
4）Comparison with Dropout and random noise
random erasing 更胜一筹
5）Comparing with data augmentation methods
单独的话，random cropping > random flipping > random erasing
三合一 1+1+1>1，猛
6）Robustness to occlusion
手动遮挡，测试下效果 random erasing 的效果
We randomly select a region of area and fill it with random values. aspect ratio [0.3, 3.33]
秀
5.2.2 Object Detection
秀
5.2.3 Person Reidentification
秀
SOTA 配合 re-ranking，结果更进一层
6 Conclusion（own） / Future work
-
arXiv-2017-11-16 挂出来的论文中了 2020 AAAI，消失的时间
-
Deep learning: Dropout, DropConnect
12中主要的Dropout方法：如何应用于DNNs，CNNs，RNNs中的数学和可视化解释
- 《A-Fast-RCNN: Hard Positive Generation via Adversary for Object Detection》（CVPR-2017）
通过训练（对特征图进行遮挡和仿射变换）提升检测网络对遮挡、形变物体的识别精度
- 《PatchShuffle Regularization》（arXiv-2017）
