Course contentsShow
← Machine Learning and Deep Learning
Lesson 955 of 3,53822. Computer Vision TasksFree lesson

R-CNN Architecture

Study the original R-CNN: region proposals, CNN feature extraction, and SVM classification pipeline.

R-CNN Architecture

What you'll learn: How the original R-CNN (Regions with CNN features) revolutionized object detection by combining region proposals with deep learning feature extraction.

The Core Innovation

R-CNN was the breakthrough that brought deep learning to object detection. Instead of trying to detect objects everywhere at once, R-CNN takes a three-stage approach that mirrors human visual attention: first identify interesting regions, then understand what's in them, finally classify precisely what you found.

Analogy: Imagine searching for specific items in a cluttered room. You'd first scan for "areas of interest" (a pile, a shelf), then look closely at each area to understand its features, and finally decide if your target item is there. That's exactly R-CNN's pipeline.

The Three-Stage Pipeline

1. Region Proposals (~2000 regions)

R-CNN uses Selective Search (a non-learning algorithm) to generate approximately 2,000 candidate bounding boxes per image. These proposals identify regions that might contain objects, drastically reducing the search space from all possible sliding windows.

2. CNN Feature Extraction

Each proposed region is:

  • Warped to a fixed size (227×227 pixels)
  • Passed through a pretrained CNN (originally AlexNet)
  • Converted into a 4096-dimensional feature vector

This transfers the power of ImageNet-pretrained features to the detection task.

3. SVM Classification

For each object class, a separate Support Vector Machine (SVM) classifier is trained on the CNN features. SVMs output class scores, and bounding box regression refines the coordinates.

Why It Mattered

R-CNN achieved 53.3% mAP on PASCAL VOC 2012—a massive jump from 35.1% with previous methods. It proved that CNNs could extract features powerful enough for detection, not just classification.

Key Takeaway: R-CNN pioneered modern object detection by decoupling region proposal from deep feature learning, using region proposals → CNN features → SVM classification to achieve breakthrough accuracy.