[Paper] Random Erasing (RE): Random Erasing Data Augmentation (Image Classification)
Improves Models for Image Classification, Object Detection & Person Re-identification
--
In this story, Random Erasing Data Augmentation (Random Erasing, RE), by Xiamen University, University of Technology Sydney, Australian National University, and Carnegie Mellon University, is shortly presented. In this paper:
- Random Erasing is proposed to randomly select a rectangle region in an image and erases its pixels with random values.
- This reduces the risk of overfitting and makes the model robust to occlusion.
- It is is complementary to commonly used data augmentation techniques such as random cropping and flipping.
This is a paper in 2020 AAAI with over 600 citations. (Sik-Ho Tsang @ Medium)
Outline
- Random Erasing (RE)
- Ablation Study
- Experimental Results
1. Random Erasing (RE)
1.1. Random Erasing (RE) Algorithm
- For an image I in a mini-batch, the probability of it undergoing Random Erasing is p.
- Random Erasing randomly selects a rectangle region Ie in an image, and erases its pixels with random values.
- The area of the image is S = W ×H. The area of erasing rectangle region is randomized as Se, where Se/S is in range specified by minimum sl and maximum sh.
- The aspect ratio re of erasing rectangle region is randomly initialized between r1 and r2.
- The size of Ie is:
- and:
- If xe+We ≤ W and ye+He ≤ H, we set the region, Ie = (xe, ye, xe +We, ye +He), as the selected rectangle region.
- With the selected erasing region Ie, each pixel in Ie is assigned to a random value in [0, 255].
- The detailed RE algorithm is as below:
1.2. Random Erasing for Image Classification and Person Re-identification
- In general, training data does not provide the location of the object. In this case, Random Erasing is performed on the whole image.
1.3. Random Erasing for Object Detection
- There are 3 schemes.
- Image-aware Random Erasing (IRE): selecting erasing region on the whole image.
- Object-aware Random Erasing (ORE): selecting erasing regions in the bounding box of each object. if there are multiple objects in the image, Random Erasing is applied on each object separately.
- Image and object-aware Random Erasing (I+ORE): selecting erasing regions in both the whole image and each object bounding box.
1.4. Comparison with Random Cropping
- Random cropping reduces the contribution of the background.
- CNN can learn the model on the presence of parts of the object instead of focusing on the whole object.
- Random Erasing retains the overall structure of the object, only occluding some parts of object. Areas are re-assigned with random values, which can be viewed as adding noise to the image.
- They can be complementary to each other.
2. Ablation Study
2.1. The impact of hyper-parameters
- Pre-Activation ResNet-18 is used as baseline.
- 3 hyperparameters to evaluate, i.e., the erasing probability p, the area ratio range of erasing region sl and sh, and the aspect ratio range of erasing region r1 and r2.
- To simplify experiment, sl is fixed to 0.02, r1 = 1/r2 and evaluate p, sh, and r1.
- p = 0.5, sh = 0.4 and r1 = 0.3 as the base setting, and alter one of them.
- When p ∈ [0.2, 0.8] and sh ∈ [0.2, 0.8], the average classification error rate is 4.48%, outperforming the baseline method (5.17%) by a large margin.
- For aspect ratio, the best result are obtained when r1 = 0.3, error rate = 4.31%, reduces the classification error rate by 0.86% compared with the baseline.
- p = 0.5, sl = 0.02, sh = 0.4, and r1 = 1/r2 = 0.3 as default settings.
2.2. Four Types of Random Values
- RE-R: Random value ranging in [0, 255].
- RE-M: mean ImageNet value.
- RE-0: 0.
- RE-255: 255.
- RE-R achieves approximately equal performance to RE-M, RE-R is chosen.
3. Experimental Results
3.1. Image Classification
- p = 0.5, sl = 0.02, sh = 0.4, and r1 = 1/r2 = 0.3.
- For CIFAR-10, random erasing improves the accuracy by 0.49% using ResNet-110.
- Random erasing obtains 3.08% error rate using WRN-28–10, which improves the accuracy by 0.72%.
- For CIFAR-100, random erasing obtains 17.73% error rate which gains 0.76% than the WRN-28-10 baseline.
- Random erasing improves WRN-28–10 from 4.01% to 3.65% in top-1 error on Fashion-MNIST.
- Random erasing consistently improves the results on all three ResNet variants on ImageNet.
3.2. Comparison with Dropout and Random Noise
- Applying Dropout or adding random noise at the image layer fails to improve the accuracy.
3.3. Comparing with Data Augmentation Methods
- RF: Random flipping, RC: Random cropping, RE: Random Erasing.
- Random Erasing and the two competing techniques are complementary. Particularly, combining these three methods achieves 4.31% error rate, a 7% improvement over the baseline without any augmentation.
3.4. Robustness to Occlusion
- The baseline performance drops quickly when increasing the occlusion level l.
- Random erasing approach achieves 56.36% error rate when the occluded area is half of the image (l = 0.5), while the baseline rapidly drops to 75.04%.
- Random Erasing improves the robustness of CNNs against occlusion.
3.5. Object Detection
- Faster R-CNN using VGG-16 backbone is used as baseline.
- For Random Erasing, p = 0.5, sl = 0.02, sh = 0.2, and r1 = 1/r2=0.3.
- The baseline got 69.1% mAP.
- The detector training with I+ORE obtains further improved in performance with 71.5% mAP.
- When using the enlarged 07+12 training set, 76.2% mAP is achieved.
3.6. Person Re-identification
- For Random Erasing, p = 0.5, sl = 0.02, sh = 0.2, and r1 = 1/r2=0.3.
- For Market-1501, Random Erasing improves the rank-1 by 3.10% and 2.67% for IDE and SVDNet with using ResNet-50.
- For DukeMTMC-reID, Random Erasing increases the rank-1 accuracy from 71.99% to 74.24% for IDE (ResNet-50) and from 76.82% to 79.31% for SVDNet (ResNet-50).
- For CUHK03, TriNet gains 8.28% and 5.0% in rank-1 accuracy when applying Random Erasing.
- This indicates that Random erasing can reduce the risk of over-fitting and improves the re-ID performance.
References
[2020 AAAI] [Random Erasing (RE)]
Random Erasing Data Augmentation
GitHub: https://github.com/zhunzhong07/Random-Erasing
Image Classification
1989–1998: [LeNet]
2012–2014: [AlexNet & CaffeNet] [Maxout] [Dropout] [NIN] [ZFNet] [SPPNet]
2015: [VGGNet] [Highway] [PReLU-Net] [STN] [DeepImage] [GoogLeNet / Inception-v1] [BN-Inception / Inception-v2]
2016: [SqueezeNet] [Inception-v3] [ResNet] [Pre-Activation ResNet] [RiR] [Stochastic Depth] [WRN] [Trimps-Soushen]
2017: [Inception-v4] [Xception] [MobileNetV1] [Shake-Shake] [Cutout] [FractalNet] [PolyNet] [ResNeXt] [DenseNet] [PyramidNet] [DRN] [DPN] [Residual Attention Network] [IGCNet / IGCV1] [Deep Roots]
2018: [RoR] [DMRNet / DFN-MR] [MSDNet] [ShuffleNet V1] [SENet] [NASNet] [MobileNetV2] [CondenseNet] [IGCV2] [IGCV3] [FishNet] [SqueezeNext] [ENAS] [PNASNet] [ShuffleNet V2] [BAM] [CBAM] [MorphNet] [NetAdapt] [mixup] [DropBlock]
2019: [ResNet-38] [AmoebaNet] [ESPNetv2] [MnasNet] [Single-Path NAS] [DARTS] [ProxylessNAS] [MobileNetV3] [FBNet] [ShakeDrop] [CutMix]
2020: [Random Erasing (RE)]