Data Augmentation
Why Augment?
Small datasets overfit — augmentation multiplies your data
With so few samples, the model memorizes training data. Validation loss diverges sharply from training loss — classic overfitting. Augmentation (orange line) narrows the gap by creating diverse variations.
Step 1 of 6: Why Augment?
Small datasets overfit — augmentation multiplies your data
With so few samples, the model memorizes training data. Validation loss diverges sharply from training loss — classic overfitting. Augmentation (orange line) narrows the gap by creating diverse variations.
Step 2 of 6: Geometric Transforms
Rotate, flip, crop — spatial invariances
Small rotations (< 30°) are the safest geometric augmentation. They preserve the object's natural orientation while adding useful variation for the model to learn invariance.
Step 3 of 6: Color & Intensity
Brightness, contrast, and color jitter
Subtle color shifts simulate natural lighting variation. The histogram barely changes, keeping pixel statistics close to the original — low-risk, high-value augmentation.
Step 4 of 6: Cutout & Mixup
Occlusion and interpolation strategies
Moderate cutout (10-30%) forces the model to attend to broader features rather than fixating on one discriminative region. Mixup blending creates soft label targets that improve calibration.
Step 5 of 6: Augmentation Probability
How often to apply each transform
Moderate probability (0.2-0.7) gives a healthy mix of original and augmented samples. Each epoch sees different variations, maximizing the effective training set diversity.
Step 6 of 6: When Augmentation Hurts
Too much augmentation degrades performance
The sweet spot: enough augmentation to teach useful invariances without distorting the data distribution. Validation accuracy peaks here, well above the no-augmentation baseline.