Object Detection

🔍 Object Detection: Teaching Computers to Find Things!

Imagine you’re playing “I Spy” with a computer. You want it to find ALL the cats in a photo AND draw boxes around each one. That’s Object Detection!

🎯 What is Object Detection?

Think about this: You look at a crowded playground photo. In 2 seconds, you spot:

• 3 kids on swings
• 1 dog running
• 2 balls on the ground

You didn’t just SEE the photo. You FOUND objects AND knew WHERE they are.

Object Detection teaches computers to do the same thing!

Two Jobs in One

Simple Example:

• Photo: A park scene
• Classification says: “park, trees, people”
• Object Detection says: “Person at box (10,20,50,80), Dog at box (100,150,60,40), Tree at box (200,10,100,300)”

📦 Bounding Box Prediction

What’s a Bounding Box?

It’s just a rectangle! Like drawing a box around something with a marker.

    ┌─────────────┐
    │    🐕       │ ← The box that says
    │   (dog)     │   "a dog is HERE!"
    └─────────────┘

The Four Magic Numbers

Every box needs 4 numbers. Think of it like giving directions:

Box = (x, y, width, height)

      x = how far from left edge
      y = how far from top edge
  width = how wide the box is
 height = how tall the box is

Real Example:

• x = 50 (50 pixels from left)
• y = 30 (30 pixels from top)
• width = 100 (box is 100 pixels wide)
• height = 80 (box is 80 pixels tall)

The computer draws a 100×80 rectangle starting at position (50, 30).

Alternative Format: Two Corners

Some systems use corner coordinates instead:

Box = (x_min, y_min, x_max, y_max)
     (top-left corner, bottom-right corner)

Both work! Just different ways to describe the same rectangle.

⚓ Anchor Boxes and NMS

The Problem: Too Many Guesses!

When a computer looks for objects, it’s like asking 1000 friends to guess where the cat is. Everyone points somewhere different!

Result: 50 boxes around one cat. We only need ONE box!

Anchor Boxes: Smart Starting Points

Instead of random guesses, we use anchor boxes—pre-made box templates.

graph TD
    A["Image Grid Cell"] --> B["Anchor 1: Tall & Thin"]
    A --> C["Anchor 2: Square"]
    A --> D["Anchor 3: Wide & Short"]
    B --> E["Good for people standing"]
    C --> F["Good for balls, faces"]
    D --> G["Good for cars, cats lying down"]

Think of it like cookie cutters:

• You have 3-5 shapes ready
• Each grid cell tries all shapes
• Pick the one that fits best!

NMS: Non-Maximum Suppression

NMS is the referee that picks the BEST box and removes duplicates.

How NMS Works (Like a Game Show):

1. Line up all boxes by confidence score (highest first)
2. Winner stays! The most confident box wins
3. Remove overlapping losers that cover the same object
4. Repeat for remaining boxes

Before NMS:          After NMS:
┌──────┐
│┌────┐│               ┌────┐
││ 🐕 ││    →→→        │ 🐕 │
│└────┘│               └────┘
└──────┘
(5 overlapping boxes)  (1 clean box!)

📐 IoU Metric: How Good is Your Box?

What is IoU?

IoU = Intersection over Union

It measures: “How much do two boxes overlap?”

Think of two paper circles on a table:

• Intersection: The part where BOTH circles cover
• Union: The total area covered by EITHER circle

IoU = (Overlap Area) ÷ (Total Area)

    ┌───────┐
    │   ┌───┼───┐
    │   │///│   │   /// = Intersection
    └───┼───┘   │
        └───────┘

IoU Score Guide

Example:

• Your predicted box: (10, 10, 50, 50)
• Actual object box: (12, 8, 48, 52)
• They overlap a lot → IoU = 0.85 ✅ Great job!

NMS uses IoU too! If two boxes have IoU > 0.5, one gets removed.

🏆 mAP Metric: The Report Card

What is mAP?

mAP = Mean Average Precision

It’s like a grade for your detector. “How good are you at finding ALL objects correctly?”

Breaking it Down

Precision: Of all boxes you drew, how many were correct?

Precision = Correct Boxes ÷ All Boxes You Made

Recall: Of all real objects, how many did you find?

Recall = Objects Found ÷ All Real Objects

Average Precision (AP): Combines precision and recall at different confidence levels

mAP: Average of AP across all object categories

Real Example

Your detector looking for cats and dogs:

• Cat AP: 0.85 (great at finding cats!)
• Dog AP: 0.75 (good at dogs)
• mAP = (0.85 + 0.75) ÷ 2 = 0.80

Your detector gets an 80% grade! 🎉

IoU Thresholds Matter

• mAP@50: Counts a detection as correct if IoU ≥ 0.5
• mAP@75: Stricter! IoU ≥ 0.75
• mAP@[50:95]: Average across multiple thresholds (hardest!)

⚡ YOLO Architecture: You Only Look Once

The Speed Champion

Before YOLO, detectors looked at images multiple times. Slow!

YOLO’s Big Idea: Look at the WHOLE image ONCE and predict EVERYTHING together!

graph TD
    A["Input Image"] --> B["Divide into Grid"]
    B --> C["Each Cell Predicts"]
    C --> D["Boxes + Classes"]
    D --> E["NMS Cleanup"]
    E --> F["Final Detections"]

How YOLO Works

Step 1: Divide image into grid (like a tic-tac-toe board, but bigger—maybe 13×13)

Step 2: Each cell predicts:

• Multiple bounding boxes
• Confidence scores
• Class probabilities (is it a cat? dog? car?)

Step 3: Combine predictions from all cells

Step 4: NMS removes duplicate boxes

YOLO Output

Each grid cell outputs:

[x, y, w, h, confidence, class1, class2, ...]
 └─ box coords ─┘    │         └─ probabilities ─┘
              "Is there an object here?"

Why YOLO is Amazing

Example: YOLO can process video in real-time—detecting objects in every frame of a live camera feed!

🔬 R-CNN Family: Accurate but Careful

The Accuracy Champions

While YOLO is fast, R-CNN family focuses on being very precise.

R-CNN (The Original)

graph TD
    A["Image"] --> B["Propose ~2000 Regions"]
    B --> C["Resize Each Region"]
    C --> D["CNN Features per Region"]
    D --> E["SVM Classifies Each"]
    E --> F["Bounding Box Refinement"]

Problem: Runs CNN 2000 times! Very slow (47 seconds per image 😴)

Fast R-CNN (Smarter)

Key improvement: Run CNN once on whole image, THEN extract features for regions.

Image → CNN → Feature Map → Extract region features → Classify

Speed: 2000× faster than R-CNN!

Faster R-CNN (Even Smarter!)

Key improvement: Use a neural network to propose regions too!

RPN (Region Proposal Network):

• Slides over feature map
• At each position, predicts “Is there an object? How big?”
• Much faster than old region proposal methods

graph TD
    A["Image"] --> B["Backbone CNN"]
    B --> C["Feature Map"]
    C --> D["RPN: Region Proposals"]
    C --> E["ROI Pooling"]
    D --> E
    E --> F["Classification + Box Refinement"]

R-CNN Family Comparison

🔺 Feature Pyramid Network (FPN)

The Problem: Big and Small Objects

Imagine finding:

• A tiny ant in a photo
• A huge elephant in the same photo

Early layers in CNNs see small details (good for ants). Late layers see big concepts (good for elephants).

Old detectors: Only used late layers. Missed small objects!

FPN’s Solution: Use ALL Layers!

graph TD
    subgraph Bottom-Up
    A["Input"] --> B["Low Level"]
    B --> C["Mid Level"]
    C --> D["High Level"]
    end

    subgraph Top-Down
    D --> E["P5"]
    E --> F["P4"]
    F --> G["P3"]
    end

    C -.->|Add| F
    B -.->|Add| G

How FPN Works

Bottom-Up Path: Normal CNN—image gets smaller, features get richer

Top-Down Path:

1. Start from smallest, richest features
2. Upsample (make bigger)
3. Add to earlier layer features
4. Now EVERY level has rich features!

Lateral Connections

The magic is in the “adding” step:

• Take high-level features (knows WHAT objects are)
• Add to low-level features (knows WHERE details are)
• Get both! Strong features at every size.

Why FPN Matters

Example: Detecting both a person and their watch in one image. The person is 500 pixels tall, the watch is 20 pixels. FPN handles both!

🎯 Putting It All Together

Modern object detectors combine these ideas:

graph TD
    A["Image"] --> B["Backbone + FPN"]
    B --> C["Anchors at Each Level"]
    C --> D["Predict Boxes + Classes"]
    D --> E["NMS"]
    E --> F["Final Detections"]
    F --> G["Evaluate with mAP"]

Quick Summary

🚀 You Did It!

You now understand how computers:

1. ✅ Find multiple objects in images
2. ✅ Draw boxes around them
3. ✅ Handle objects of different sizes
4. ✅ Choose the best predictions

These same techniques power:

• Self-driving cars spotting pedestrians
• Phones detecting faces
• Security cameras finding unusual activity
• Robots picking up objects

You’re ready to build your own object detector! 🎉