Course contentsShow
← Machine Learning and Deep Learning
Lesson 956 of 3,53822. Computer Vision TasksFree lesson

Fast R-CNN Improvements

Learn how Fast R-CNN speeds up training and inference with RoI pooling and end-to-end training.

Fast R-CNN Improvements

What You'll Learn

You'll discover how Fast R-CNN dramatically accelerated object detection by introducing RoI pooling and end-to-end training, solving critical bottlenecks in the original R-CNN.

The Problem with R-CNN

Remember R-CNN? It was powerful but painfully slow. Why? Because it extracted ~2000 region proposals per image, ran each through a CNN separately, and trained multiple disconnected components (CNN, SVM classifier, bounding box regressor). This meant processing the same image thousands of times—incredibly wasteful!

Fast R-CNN's Key Innovations

Single-Pass Feature Extraction

Instead of running the CNN 2000 times, Fast R-CNN processes the entire image once through convolutional layers to create a feature map. Think of it like photographing a crowd once, rather than taking separate photos of each person.

Region of Interest (RoI) Pooling

Here's the clever part: RoI pooling extracts fixed-size feature vectors from variable-sized region proposals on the shared feature map. It divides each region into a fixed grid (say 7×7) and max-pools within each cell, producing uniform output regardless of input region size. This allows batch processing of all proposals simultaneously.

End-to-End Training

Fast R-CNN unified the pipeline! Instead of separately training a CNN, classifier, and bounding box regressor, everything trains together with a multi-task loss: classification loss + localization loss. Gradients flow through the entire network, making training more efficient and effective.

Speed Improvements

  • Training: ~9× faster than R-CNN
  • Inference: ~140× faster than R-CNN
  • Both classification and bounding box regression happen in one forward pass

Key Takeaway: Fast R-CNN revolutionized object detection by sharing convolutional computations across proposals through RoI pooling and unifying training with an end-to-end multi-task loss, achieving dramatic speedups while maintaining accuracy.