random erasing

**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.