Random Erasing is a data augmentation technique that randomly selects a rectangular region in the image and replaces its pixels with random values or a fixed value — similar to Cutout but with random aspect ratios and fill values for greater variety.

How Does Random Erasing Work?

• Probability: Apply erasing with probability $p$ (typically 0.5).
• Area: Erase a region with area ratio $s in [0.02, 0.4]$ of the total image.
• Aspect Ratio: Random aspect ratio $r in [0.3, 3.3]$ for the erased region.
• Fill: Replace with random pixel values, zeros, or ImageNet mean values.
• Paper: Zhong et al. (2020).

Why It Matters

• More Varied Than Cutout: Random aspect ratios and fill values create more diverse occlusion patterns.
• Person Re-ID: Particularly effective for person re-identification where occlusion is common.
• Stacking: Can be combined with other augmentations (Mixup, CutMix) for additive benefits.

Random Erasing is Cutout with variety — randomly occluding rectangular regions with flexible shapes and fill patterns.