Data augmentation for vision
Expanding training images with flips, crops, color jitter, and related transforms so models generalize better.
What it is
Expanding training images with flips, crops, color jitter, and related transforms so models generalize better.
Why it matters
Often the cheapest accuracy win—and a regularization method (Course 05).
How it works (plain)
Apply label-preserving transforms at train time. Don’t flip digits that change meaning (6/9 hazards). Match test domain.
Try it
List 5 safe augmentations for product photos and 2 unsafe ones.
Myths
- ⚠️ Myth: More augmentation always helps.
- ✓ Reality: Too strong destroys signal or changes labels.
Sources
- Course 05 regularization; CNNs; transfer learning
- fast.ai: https://www.fast.ai/ ↗