Why augmentation multiplies training data in Computer Vision - Why Metrics Matter

Metrics & Evaluation - Why augmentation multiplies training data

Which metric matters for this concept and WHY

When using data augmentation in computer vision, the key metric to watch is generalization performance, often measured by validation accuracy or validation loss. Augmentation creates new, varied images from existing ones, helping the model learn to recognize objects under different conditions. This reduces overfitting and improves how well the model works on new, unseen images. So, metrics that show how well the model performs on data it has never seen before are the most important.

Confusion matrix or equivalent visualization (ASCII)

Confusion Matrix Example (Augmented Data Model):

          Predicted
          Cat   Dog
Actual Cat  45    5
       Dog   4    46

Total samples: 100

- True Positives (TP): 45 (correctly predicted Cat)
- True Negatives (TN): 46 (correctly predicted Dog)
- False Positives (FP): 4 (Dog predicted as Cat)
- False Negatives (FN): 5 (Cat predicted as Dog)

Precision = TP / (TP + FP) = 45 / (45 + 4) = 0.918
Recall = TP / (TP + FN) = 45 / (45 + 5) = 0.9
F1 Score = 2 * (Precision * Recall) / (Precision + Recall) ≈ 0.909

Precision vs Recall tradeoff with concrete examples

Imagine a model trained with and without augmentation:

• Without augmentation: The model sees fewer variations, so it might memorize training images but fail on new ones. This can cause high precision but low recall because it only confidently predicts what it has seen.
• With augmentation: The model learns to recognize objects in many forms (rotated, brightened, flipped). This usually improves recall because it finds more true positives, even if precision slightly drops due to more challenging cases.

For example, in a dog vs cat classifier, augmentation helps the model spot a dog even if the photo is blurry or rotated, increasing recall.

What "good" vs "bad" metric values look like for this use case

Good metrics:

• Validation accuracy above 85% on varied images
• Balanced precision and recall around 90%
• F1 score close to precision and recall, showing consistent performance

Bad metrics:

• High training accuracy but low validation accuracy (overfitting)
• Very high precision but very low recall (model misses many true cases)
• Validation accuracy below 70%, showing poor generalization

Metrics pitfalls (accuracy paradox, data leakage, overfitting indicators)

• Accuracy paradox: High accuracy can be misleading if classes are imbalanced. For example, if 90% of images are cats, predicting all cats yields 90% accuracy but no real learning.
• Data leakage: Augmented images too similar to training images in validation can inflate metrics falsely.
• Overfitting indicators: Large gap between training and validation accuracy means the model memorizes training data but fails to generalize.

Self-check question

Your model trained with augmentation has 98% training accuracy but only 12% recall on the dog class in validation. Is it good for production? Why or why not?

Answer: No, it is not good. The very low recall means the model misses most dogs, which is critical if detecting dogs is important. Despite high training accuracy, the model fails to generalize and recognize dogs in new images. This suggests overfitting or ineffective augmentation.

Key Result

Data augmentation improves model generalization by increasing validation accuracy and recall, reducing overfitting.

Practice

1. Why does image augmentation multiply the training data in computer vision?

easy

A. Because it creates new images by changing existing ones

B. Because it deletes some images to reduce overfitting

C. Because it compresses images to save space

D. Because it labels images automatically

Solution

1. Step 1: Understand augmentation purpose

   Augmentation creates new images by applying changes like rotation or flipping to existing images.

2. Step 2: Connect augmentation to data size

   These new images increase the total number of training samples, effectively multiplying the dataset size.

3. Final Answer:

   Because it creates new images by changing existing ones -> Option A

4. Quick Check:

   Augmentation = new image versions [OK]

Hint: Augmentation makes new images from old ones [OK]

Common Mistakes:

• Thinking augmentation deletes images
• Confusing augmentation with compression
• Believing augmentation automatically labels data

2. Which of the following Python code snippets correctly applies horizontal flip augmentation using torchvision transforms?

easy

A. transform = transforms.FlipHorizontal()

B. transform = transforms.HorizontalFlip()

C. transform = transforms.RandomHorizontalFlip(p=1.0)

D. transform = transforms.RandomFlip(direction='horizontal')

Solution

1. Step 1: Recall torchvision syntax

   In torchvision, horizontal flip is done with RandomHorizontalFlip and a probability parameter.

2. Step 2: Check options for correct class and parameters

   Only transform = transforms.RandomHorizontalFlip(p=1.0) uses the correct class name and parameter format.

3. Final Answer:

   transform = transforms.RandomHorizontalFlip(p=1.0) -> Option C

4. Quick Check:

   Correct torchvision flip syntax = transform = transforms.RandomHorizontalFlip(p=1.0) [OK]

Hint: Look for RandomHorizontalFlip with p parameter [OK]

Common Mistakes:

• Using incorrect class names like HorizontalFlip
• Missing the probability parameter p
• Assuming FlipHorizontal exists in torchvision

3. Given this code snippet applying augmentation, what is the total number of images after augmentation if the original dataset has 100 images and each image is augmented with 3 different transformations?

medium

A. 300

B. 400

C. 100

D. 600

Solution

1. Step 1: Calculate augmented images per original

   Each original image produces 3 new augmented images plus the original itself, totaling 4 images per original.

2. Step 2: Multiply by original dataset size

   100 original images * 4 images each = 400 images total after augmentation.

3. Final Answer:

   400 -> Option B

4. Quick Check:

   (3 augmentations + 1 original) * 100 = 400 [OK]

Hint: Add original + augmentations, then multiply by dataset size [OK]

Common Mistakes:

• Counting only augmented images, ignoring originals
• Multiplying by 3 instead of 4
• Adding instead of multiplying

4. Identify the error in this augmentation code snippet and select the fix:

from torchvision import transforms
transform = transforms.Compose([
    transforms.RandomRotation(30),
    transforms.RandomFlip(horizontal=true)
])

medium

A. Replace RandomFlip with RandomHorizontalFlip

B. Change RandomRotation(30) to RandomRotation(degrees=30)

C. Add p=0.5 to RandomFlip

D. No error, code is correct

Solution

1. Step 1: Check torchvision augmentation classes

   There is no RandomFlip class in torchvision; horizontal flip uses RandomHorizontalFlip.

2. Step 2: Identify correct fix

   Replacing RandomFlip with RandomHorizontalFlip fixes the error; other options are incorrect or unnecessary.

3. Final Answer:

   Replace RandomFlip with RandomHorizontalFlip -> Option A

4. Quick Check:

   Correct flip class = RandomHorizontalFlip [OK]

Hint: Use exact torchvision class names for flips [OK]

Common Mistakes:

• Assuming RandomFlip exists
• Changing rotation syntax unnecessarily
• Adding parameters to non-existent classes

5. You have a small dataset of 50 images. You want to multiply your training data by 5 times using augmentation. Which combination of augmentations is best to achieve this without losing original images?

hard

A. Apply 5 augmentations but discard original images

B. Only keep original images and train longer

C. Apply 2 augmentations and duplicate images twice

D. Apply 4 different augmentations per image plus keep original images

Solution

1. Step 1: Understand multiplication goal

   To multiply data by 5 times, total images after augmentation should be 50 * 5 = 250.

2. Step 2: Calculate images per original

   Keeping original plus 4 augmentations per image gives 5 images per original, totaling 250 images.

3. Step 3: Evaluate options

   Apply 4 different augmentations per image plus keep original images achieves the goal without losing original images; others either discard originals or don't reach 5x.

4. Final Answer:

   Apply 4 different augmentations per image plus keep original images -> Option D

5. Quick Check:

   Original + 4 augmentations = 5x data [OK]

Hint: Keep originals + 4 augmentations for 5x data [OK]

Common Mistakes:

• Discarding original images reduces data size
• Duplicating images without augmentation doesn't add variety
• Training longer without augmentation doesn't multiply data