[Paper Review] Random Erasing Data Augmentation

Zhun Zhong, Liang Zheng+3 more|arXiv (Cornell University)|Aug 16, 2017

Video Surveillance and Tracking Methods36 references748 citations

TL;DR

Random Erasing is a parameter-free data augmentation technique that occludes random rectangular regions in training images, improving generalization to occlusion across classification, detection, and re-ID tasks.

ABSTRACT

In this paper, we introduce Random Erasing, a new data augmentation method for training the convolutional neural network (CNN). In training, Random Erasing randomly selects a rectangle region in an image and erases its pixels with random values. In this process, training images with various levels of occlusion are generated, which reduces the risk of over-fitting and makes the model robust to occlusion. Random Erasing is parameter learning free, easy to implement, and can be integrated with most of the CNN-based recognition models. Albeit simple, Random Erasing is complementary to commonly used data augmentation techniques such as random cropping and flipping, and yields consistent improvement over strong baselines in image classification, object detection and person re-identification. Code is available at: https://github.com/zhunzhong07/Random-Erasing.

Motivation & Objective

• Motivate generalization improvements for CNNs by simulating occlusion.
• Propose a simple, parameter-free augmentation that can be integrated with existing CNNs.
• Demonstrate effectiveness across image classification, object detection, and person re-identification.
• Show robustness to occlusion and complementarity with other augmentations.

Proposed method

• Randomly select a rectangle region in an image with probability p during training.
• Erase the selected region by filling it with random pixel values (or ImageNet mean)
• Adjust erasing region size via area s_l–s_h and aspect ratio r_1–r_2 to generate varying occlusion levels
• Apply Random Erasing to image classification and person re-ID by occluding regions in the full image.
• For object detection, apply three schemes: image-aware (IRE), object-aware (ORE), and image+object-aware (I+ORE)
• Compare Random Erasing with random cropping and cropping-based augmentations to study complementarities.
• No additional parameter learning is required; it is lightweight and easily integrated.

Experimental results

Research questions

• RQ1 Does Random Erasing improve generalization of CNNs under occlusion across multiple vision tasks?
• RQ2 How do erasing probability, area, and aspect ratio affect performance?
• RQ3 Is Random Erasing complementary to other augmentations like random cropping and flipping?
• RQ4 Can the method be adapted for object detection with image/object-aware schemes?
• RQ5 Does the technique provide robustness against occluded test examples?],
• RQ6 key_findingsArgumentList_minimal_1? hmm:

Key findings

• Improves accuracy across CIFAR-10, CIFAR-100, and Fashion-MNIST over baselines for various architectures (e.g., ResNet variants, WRN-28-10).
• On CIFAR-10 with ResNet-110 and WRN-28-10, Random Erasing yields notable reductions in test error (e.g., WRN-28-10 achieves 3.08% error).
• Random Erasing outperforms baseline and competes with, and complements, existing augmentations like random cropping and flipping.
• In object detection, applying IRE, ORE, or I+ORE during Fast-RCNN training yields mAP improvements (e.g., I+ORE reaching 71.5% mAP on VOC07).
• In person re-ID, Random Erasing consistently improves rank-1 and mAP across Market-1501, DukeMTMC-reID, and CUHK03 baselines.
• Random Erasing increases robustness to occlusion, with CIFAR-10 occlusion experiments showing slower performance degradation compared to baseline.