Computer Vision Datasets and Annotation

# CHAPTER 15

Computer Vision Datasets and Annotation #

1. Introduction #

2. Learning Objectives #

• Explain why custom datasets are required for enterprise CV tasks.

• Define Image Annotation and Bounding Box labeling.

• Identify popular tools used by data scientists to label images.

• Understand the concept of Data Augmentation to artificially expand datasets.

3. Beginner-Friendly Explanation #

4. The Annotation Process (Object Detection) #

5. Types of Image Annotation #

• Image Classification: The easiest. You just put 500 photos of apples into a folder named "Apples", and 500 photos of oranges into a folder named "Oranges."

• Object Detection (Bounding Boxes): Drawing rectangles around objects. Used for cars, people, and specific items.

• Image Segmentation (Polygons): The hardest. Instead of a box, humans must manually click and trace the *exact* pixel outline of the object. This is required for self-driving cars so the AI knows exactly where the road ends and the sidewalk begins.

6. Popular Annotation Tools #

• CVAT (Computer Vision Annotation Tool): A free, open-source web tool built by Intel. It is the industry standard for collaborative labeling.

• LabelImg: A simple, free desktop app that lets you quickly draw bounding boxes and export them in YOLO format.

• Roboflow: A modern, cloud-based platform that helps teams manage, label, and augment computer vision datasets seamlessly.

7. Data Augmentation (Cheating the System) #

1. 1. Flips it horizontally.

1. 2. Increases the brightness by 20%.

1. 3. Adds 10% digital noise/blur.

1. 4. Rotates it 15 degrees.

8. Python Example: Basic Data Augmentation #

import cv2
import numpy as np

img = cv2.imread("training_dog.jpg")

# Augmentation 1: Flip Horizontally (1 means horizontal)
flipped_img = cv2.flip(img, 1)

# Augmentation 2: Add random noise
noise = np.random.normal(0, 25, img.shape).astype(np.uint8)
noisy_img = cv2.add(img, noise)

# Save the new training images!
cv2.imwrite("aug_flipped.jpg", flipped_img)
cv2.imwrite("aug_noisy.jpg", noisy_img)

9. Mini Project #

10. Best Practices #

• Tight Bounding Boxes: When labeling data, the box must perfectly touch the outermost pixels of the object. If your boxes are loose and include a lot of background grass, the AI will think "grass" is part of the definition of a "Dog."

11. Common Mistakes #

• Imbalanced Datasets: If your dataset has 9,000 labeled photos of Cars, but only 100 labeled photos of Motorcycles, the AI will be horribly biased. It will likely guess "Car" every time it sees a Motorcycle because it is statistically safer. Always balance your classes!

12. Exercises #

1. 1. Explain why Image Segmentation (Polygon labeling) is significantly more expensive and time-consuming for companies than standard Bounding Box labeling.

13. MCQs with Answers #

14. Interview Questions #

• Q: Describe the difference in data preparation required for an Image Classification task versus an Object Detection task.

• Q: What is Data Augmentation, and why is it a mandatory step when training a Deep Learning vision model on a small dataset?

15. FAQs #

16. Summary #

17. Next Chapter Recommendation #