Cutout

Cutout is a data augmentation technique that randomly masks (zeroes out) a square region of the input image — forcing the model to learn from partial information and preventing over-reliance on any single region of the image.

How Does Cutout Work?

• Random Position: Select a random center position $(x, y)$ in the image.
• Mask: Zero out a square patch of size $L \times L$ centered at $(x, y)$.
• Boundary: The mask can extend beyond the image boundary (partial occlusion is still applied).
• Typical Size: $L = 16$ for CIFAR-10 (32×32 images), scaling up for larger images.
• Paper: DeVries & Taylor (2017).

Why It Matters

• Robustness: Teaches the model to classify using any visible part of the object, not just the most discriminative region.
• Occlusion Handling: Simulates real-world partial occlusion scenarios.
• Simple & Effective: Consistently improves accuracy by 0.5-1.0% on CIFAR and ImageNet with no tuning.

Cutout is learning with missing information — randomly hiding parts of the image to create a more robust feature extractor.