Step by Step
F
Flip — mirroring the image
Horizontally flipping an image creates a new, valid training example, since most objects (like a cat) remain recognizable when mirrored.
Example: a single cat photo horizontally flipped, creating a second valid training example from the exact same underlying photo.
R
Rotate — changing orientation
Rotating an image by some amount creates another valid training variation, teaching the model that an object's identity doesn't depend on its exact orientation.
Example: rotating a cat photo by 15 degrees, teaching the model that a slightly tilted cat is still recognizably a cat.
C
Crop — changing framing and scale
Randomly cropping an image creates variations in framing and effective scale, teaching the model to recognize objects regardless of exactly how they're framed or how much of the surrounding context is visible.
Example: randomly cropping a cat photo to show just the upper half of the cat, teaching the model to still recognize "cat" even from this partial, differently-scaled view.
B
Brighten — and other transformations, adding noise
Adjusting brightness/contrast and adding noise are additional common augmentation techniques, and a single original photo can become roughly 20 distinct training examples once all these transformations are combined and applied.
Example: a single original cat photo, once flipped, rotated, cropped, brightened, and noise-adjusted in various combinations, effectively becoming roughly 20 distinct training examples.
Applied Walkthrough
1
A team has only a limited number of real photos to train a cat-recognition model, far fewer than would ideally be needed.
2
They apply data augmentation: flipping, rotating, cropping, and adjusting brightness on each original photo, generating many new valid training variations from the same underlying images.
3
A single original cat photo can become roughly 20 distinct training examples once all these transformations are combined and applied during training.
4
This is exactly why modern CV models can generalize well even with limited real training data — augmentation happens on-the-fly during training, effectively teaching the model that a cat is still a cat whether flipped, rotated, or partially cropped.
Exam Application
Exams test whether you can name common augmentation techniques (flip, rotation, crop, brightness/contrast, noise) and whether you understand why these transformations improve model robustness — teaching invariance to variations that shouldn't change the correct prediction.
⚠ Common Trap
The most common trap is assuming data augmentation happens once, as a separate preprocessing step, permanently expanding a stored dataset. In practice, augmentation typically happens on-the-fly during training, generating different random transformations of the same original images each time they're used, rather than pre-generating and permanently storing a fixed, expanded dataset.
✓ Quick Self-Check
1. Why does flipping an image horizontally create a valid new training example?
Because most objects (like a cat) remain recognizable and correctly labeled when mirrored.
Tap to reveal / hide
2. What does rotating training images teach the model?
That an object's identity doesn't depend on its exact orientation.
Tap to reveal / hide
3. What does randomly cropping training images teach the model?
To recognize objects regardless of exactly how they're framed or how much surrounding context is visible.
Tap to reveal / hide
4. Roughly how many distinct training examples can a single original photo become through combined augmentation?
Roughly 20.
Tap to reveal / hide
5. Does data augmentation typically happen once as a preprocessing step, or on-the-fly during training?
On-the-fly during training, generating different random transformations each time the image is used, rather than a one-time preprocessing step.
Tap to reveal / hide