10-R-CNN

Define the original content for each file to ensure accuracy

📑 Comprehensive GitHub README about the evolution of the R-CNN family.

📋 Table of Contents

• Introduction
• Object Detection Pipeline
• Sliding Window
• Selective Search
• R-CNN
• Fast R-CNN
• Faster R-CNN
• Region Proposal Network (RPN)
• Comparison Tables
• Advantages & Disadvantages
• References

💡 Introduction

Object detection aims to classify and localize objects simultaneously.

🎯 Classification vs Localization vs Detection

Task	Output
Classification	Class
Localization	Class + Bounding Box
Detection	Multiple Classes + Multiple Bounding Boxes

🪟 Sliding Window

Traditional methods evaluate thousands of windows across an image, making them computationally expensive.

🔍 Selective Search

Selective Search groups similar regions using hierarchical segmentation to generate approximately 2,000 candidate object regions.

✅ Pros

• No training required
• High recall

❌ Cons

• Slow
• Hand-crafted algorithm
• Not end-to-end

🧠 R-CNN (2014)

Pipeline:

1. Selective Search
2. Warp each proposal
3. CNN feature extraction
4. SVM classification
5. Bounding-box regression

🌟 Advantages

• Huge accuracy improvement over traditional methods.

⚠️ Disadvantages

• Very slow
• Multi-stage training
• Large disk storage for extracted features

⚡ Fast R-CNN (2015)

Improvements:

• CNN runs once on the whole image.
• ROI Pooling extracts proposal features.
• Joint classification and box regression.

📥 ROI Pooling

Converts proposals with different sizes into fixed-size feature maps.

⏳ Remaining Bottleneck

Selective Search is still required.

🚀 Faster R-CNN (2015)

Major innovation: Selective Search is replaced by a Region Proposal Network (RPN).

🏗️ Region Proposal Network

The RPN slides a small network over the feature map and predicts:

• Objectness score
• Bounding box offsets

⚓ Anchor Boxes

Multiple anchors with different scales and aspect ratios are evaluated at each location.

⏩ Why Faster?

The backbone CNN is shared between proposal generation and detection.

📈 Evolution

graph LR
A[Sliding Window]
-->B[Selective Search]
-->C[R-CNN]
-->D[Fast R-CNN]
-->E[Faster R-CNN]

Loading