Lesson 32 • Advanced

Object Detection: YOLO, SSD & Faster R-CNN

Detect and localise objects in images — learn IoU, Non-Maximum Suppression, and how YOLO processes entire images in one pass.

✅ What You'll Learn

• IoU: measuring bounding box overlap
• Non-Maximum Suppression to remove duplicate detections
• YOLO: grid-based single-shot detection
• 1-stage vs 2-stage detectors and when to use each

🎯 Finding Objects in Images

🎯 Real-World Analogy: Classification is like asking "Is there a cat in this photo?" Object detection is like asking "Where are ALL the cats, dogs, and cars in this photo, and draw a box around each one." It's the difference between a yes/no question and a treasure hunt with a marker pen.

Object detection adds localisation to classification. The model must predict: (1) what objects are present, (2) where they are (bounding boxes), and (3) how confident it is. Modern detectors like YOLOv8 can process 100+ frames per second — fast enough for real-time video.

Try It: IoU & NMS

Calculate bounding box overlap and remove duplicate detections

Try it Yourself »

Python

import numpy as np

# Intersection over Union (IoU): The Core Detection Metric
# Measures how well a predicted box overlaps with the ground truth

def compute_iou(box1, box2):
    """
    Compute IoU between two bounding boxes
    Each box: [x1, y1, x2, y2] (top-left and bottom-right corners)
    """
    # Intersection coordinates
    x1 = max(box1[0], box2[0])
    y1 = max(box1[1], box2[1])
    x2 = min(box1[2], box2[2])
    y2 = min(box1[3], box2[3])
    
    # Intersection area
    intersecti
...

Try It: YOLO Architecture

See how YOLO divides images into grid cells for detection

Try it Yourself »

Python

import numpy as np

# YOLO: You Only Look Once — Real-Time Object Detection
# Process the ENTIRE image in one forward pass

np.random.seed(42)

def simulate_yolo_grid(image_size, grid_size, n_boxes, n_classes):
    """Simulate YOLO's grid-based detection"""
    cell_size = image_size // grid_size
    
    print(f"=== YOLO Architecture (v8) ===")
    print()
    print(f"Image: {image_size}x{image_size}")
    print(f"Grid:  {grid_size}x{grid_size} = {grid_size**2} cells")
    print(f"Each cell pre
...

⚠️ Common Mistake: Training object detectors without enough anchor box variety. If your objects are very tall/thin or very wide, the default anchors won't match well. Use k-means clustering on your dataset's bounding boxes to generate custom anchors.

💡 Pro Tip: For quick prototyping, use Ultralytics YOLOv8 — it's 5 lines of Python to train on custom data. For production, consider ONNX export for 2-3× faster inference. If accuracy matters more than speed, use Faster R-CNN with a ResNet-101 backbone.

📋 Quick Reference

Concept	What It Does	Threshold
IoU	Measures box overlap	≥0.5 for match
NMS	Removes duplicate boxes	IoU > 0.5 suppressed
Confidence	Object presence score	≥0.25 typically
mAP	Mean Average Precision	Higher = better
FPS	Frames per second	≥30 for real-time

🎉 Lesson Complete!

You can now detect and localise objects in images! Next, learn semantic segmentation — labelling every single pixel.