【Random Erasing】《Random Erasing Data Augmentation》

AAAI-2020

---

---

1 Background and Motivation

For promotion CNN The generalization ability of the model , promote CNN The identification ability of the model to deal with occluded scenes , A data augmentation method is proposed ,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

Put forward Random Erasing Data Augmentation, And random cropping,random filpping complementary ,not require any extra parameter learning, In the classification / testing / reID Good results in tasks

4 Method

1） Algorithm flow

$x_e$ and $y_e$ It's the center point ,$W_e$ and $H_e$ It's length and width

Random center point , Random aspect ratio , Random area , Random pixel filling value

2） Classification and ReID Application on

Simple and crude , Full picture range random

3） Application on target detection

Full picture range , Target range , Full picture + Target range

4） and Random cropping The difference between

random cropping, It can reduce the interference of the background ,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

Combined, the samples are more abundant

5 Experiments

5.1 Datasets and Metrics

1） Data sets

• classification
  CIFAR-10
  CIFAR-100
  Fashion-MNIST

• testing
  PASCAL VOC 2007

• ReID
  Market-1501
  DukeMTMC-reID
  CUHK03

2） The evaluation index

classification ,top-1 error rates,“mean std” based on 5 runs

testing ,mAP

ReID,rank-1,mAP

5.2 Experiments

5.2.1 Image Classification

1）Classification accuracy on different datasets

$p=0.5$, $s_l=0.02$, $s_h=0.4$, and $r1=1/r2=0.3$

2）The impact of hyper-parameters

fix $s_l$ to 0.02, $r1=1/r2$ and evaluate $p$, $s_h$, and $r1$

We set $p=0.5$, $s_h=0.4$ and $r1=0.3$ as the base setting. When evaluating one of the parameters, we fixed the other two parameters

All ratio Baseline（ No, random erasing） To better effect ！

3）Four types of random values for erasing

Random filling value and filling ImageNet Of mean The effect is almost the same [125, 122,114]（ I understand 114 The source of , The original gray color is ImageNet The average of ）, Better than filling 0 and 255

4）Comparison with Dropout and random noise

random erasing better

5）Comparing with data augmentation methods

Alone ,random cropping > random flipping > random erasing

Three in one 1+1+1>1, Fierce

6）Robustness to occlusion

Manually block , Test the effect random erasing The effect of

We randomly select a region of area and fill it with random values. aspect ratio [0.3, 3.33]

Show

5.2.2 Object Detection

Show

5.2.3 Person Reidentification

Show

SOTA coordination re-ranking, The result goes further

6 Conclusion（own） / Future work

• arXiv-2017-11-16 In the paper that hangs out 2020 AAAI, Time to disappear

• Deep learning: Dropout, DropConnect
  12 Main Dropout Method ： How to apply to DNNs,CNNs,RNNs Mathematical and visual interpretation in

• 《A-Fast-RCNN: Hard Positive Generation via Adversary for Object Detection》（CVPR-2017）

Through training （ Occlude and affine transform the feature map ） Improve the detection of network occlusion 、 Recognition accuracy of deformed objects

• 《PatchShuffle Regularization》（arXiv-2017）