Dataset bias in vision in Computer Vision - Model Pipeline Trace

Model Pipeline - Dataset bias in vision

This pipeline shows how dataset bias in vision can affect model training and predictions. It demonstrates how biased data leads to skewed learning and poor generalization.

Data Flow - 5 Stages

1Raw Image Dataset

1000 images x 64x64 pixels x 3 channels→Collect images mostly of cats in indoor settings→1000 images x 64x64 pixels x 3 channels

Image of a cat sitting on a couch indoors

↓

2Data Preprocessing

1000 images x 64x64 pixels x 3 channels→Resize images to 32x32 pixels and normalize pixel values→1000 images x 32x32 pixels x 3 channels

Resized and normalized cat image

↓

3Train/Test Split

1000 images x 32x32 pixels x 3 channels→Split dataset into 800 training and 200 testing images→800 training images x 32x32 pixels x 3 channels, 200 testing images x 32x32 pixels x 3 channels

Training set mostly indoor cat images, test set includes some outdoor cat images

↓

4Model Training

800 training images x 32x32 pixels x 3 channels→Train CNN to classify cat images→Trained CNN model

Model learns mostly indoor cat features

↓

5Model Evaluation

200 testing images x 32x32 pixels x 3 channels→Evaluate model accuracy on test images→Accuracy score (percentage)

High accuracy on indoor cats, low accuracy on outdoor cats

Training Trace - Epoch by Epoch

Loss
1.2 |****
0.9 |***
0.7 |**
0.6 |*
0.55|*
    +------------
     Epochs 1-5

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.2	0.45	Model starts learning basic features
2	0.9	0.60	Model improves recognizing indoor cats
3	0.7	0.72	Model fits well to biased training data
4	0.6	0.78	Loss decreases steadily, accuracy increases
5	0.55	0.82	Model converges on training data with bias

Prediction Trace - 5 Layers

Layer 1: Input Image

Layer 2: Convolutional Layer

Layer 3: Pooling Layer

Layer 4: Fully Connected Layer

Layer 5: Output Layer with Softmax

Model Quiz - 3 Questions

What is a main cause of dataset bias in this vision model?

A Most training images show cats indoors

B Images are resized to 32x32 pixels

C The model uses ReLU activation

D Test images are fewer than training images

Key Insight

Dataset bias in vision models causes them to learn features mostly from the dominant data type, reducing accuracy on underrepresented cases. Careful dataset design and diverse data collection are essential to build fair and robust vision models.

Practice

1. What does dataset bias in computer vision mean?

easy

A. The data does not fairly represent all types of images or cases

B. The model always predicts perfectly on all images

C. The dataset is too large to process

D. The images are all black and white

Solution

1. Step 1: Understand dataset bias meaning

   Dataset bias means the data used to train a model does not cover all possible cases fairly.

2. Step 2: Compare options to definition

   Only The data does not fairly represent all types of images or cases describes this correctly. Other options describe unrelated issues.

3. Final Answer:

   The data does not fairly represent all types of images or cases -> Option A

4. Quick Check:

   Dataset bias = unfair data representation [OK]

Hint: Bias means data is not fair or balanced [OK]

Common Mistakes:

• Thinking bias means model is perfect
• Confusing bias with dataset size
• Assuming bias means image color

2. Which of the following is the correct way to check for dataset bias in a vision dataset using Python?

easy

A. Use random.shuffle(dataset) to fix bias

B. Use value_counts() on labels to see class distribution

C. Use len(dataset) without checking labels

D. Use print(dataset) only

Solution

1. Step 1: Identify method to check bias

   Checking class distribution with value_counts() helps find imbalance in labels.

2. Step 2: Evaluate other options

   Printing dataset or length alone doesn't show bias. Shuffling data doesn't check bias.

3. Final Answer:

   Use value_counts() on labels to see class distribution -> Option B

4. Quick Check:

   Check label counts = value_counts() [OK]

Hint: Check label counts to find bias [OK]

Common Mistakes:

• Only printing dataset without analysis
• Assuming dataset length shows bias
• Thinking shuffling fixes bias

3. Given this Python code snippet analyzing a vision dataset labels:

import pandas as pd
labels = ['cat', 'dog', 'cat', 'cat', 'dog', 'bird']
label_counts = pd.Series(labels).value_counts()
print(label_counts)

What is the output?

medium

A. bird 3
cat 2
dog 1

B. cat 2
dog 3
bird 1

C. cat 3
dog 2
bird 1

D. dog 3
cat 3
bird 3

Solution

1. Step 1: Count occurrences of each label

   Labels list has 'cat' 3 times, 'dog' 2 times, and 'bird' 1 time.

2. Step 2: Understand value_counts output

   value_counts() returns counts sorted descending by default.

3. Final Answer:

   cat 3\ndog 2\nbird 1 -> Option C

4. Quick Check:

   Count labels correctly = cat:3, dog:2, bird:1 [OK]

Hint: Count each label frequency carefully [OK]

Common Mistakes:

• Mixing counts of dog and cat
• Assuming alphabetical order instead of count order
• Miscounting occurrences

4. You have this code to check dataset bias:

labels = ['car', 'car', 'truck', 'car', 'truck']
counts = {}
for label in labels:
    counts[label] = counts.get(label, 0) + 1
print(counts)

But the output is {}. What is the likely error?

medium

A. The 'get' method is used incorrectly with wrong parameters

B. The print statement is outside the loop and missing indentation

C. The code is correct; output should be {'car': 3, 'truck': 2}

D. The dictionary 'counts' was reinitialized inside the loop

Solution

1. Step 1: Analyze code behavior

   If 'counts' is reset inside the loop, it will be empty after loop ends.

2. Step 2: Identify cause of empty output

   Reinitializing 'counts' inside loop clears previous counts, causing empty dict at print.

3. Final Answer:

   The dictionary 'counts' was reinitialized inside the loop -> Option D

4. Quick Check:

   Resetting dict inside loop empties counts [OK]

Hint: Check if dict is reset inside loop [OK]

Common Mistakes:

• Thinking print indentation causes empty output
• Assuming get() method is wrong
• Ignoring variable scope inside loop

5. You have a vision dataset with 90% images of cats and 10% dogs. Which method best reduces dataset bias to improve model fairness?

hard

A. Collect more dog images to balance classes

B. Remove all cat images to keep only dogs

C. Train model only on cat images

D. Ignore class imbalance and train as is

Solution

1. Step 1: Understand dataset imbalance problem

   Having 90% cats and 10% dogs causes bias favoring cats.

2. Step 2: Choose method to fix bias

   Collecting more dog images balances classes, reducing bias and improving fairness.

3. Final Answer:

   Collect more dog images to balance classes -> Option A

4. Quick Check:

   Balance classes by adding data [OK]

Hint: Balance classes by adding underrepresented data [OK]

Common Mistakes:

• Removing majority class loses useful data
• Training on one class ignores others
• Ignoring imbalance causes biased model