$\color{black}\rule{365px}{3px}$

Link to Paper:

“Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks” - 2016

arxiv.org

Table of Contents

1. Introduction

$\color{black}\rule{365px}{3px}$

Motivation

The region proposal step still consumes as much running time as the detection network

Significance of the paper

• New region proposal method, Region Proposal Network(RPN), is introduced:
  • R-CNN and Fast R-CNN used Selective Search implemented for CPU.
  • New approach, RPN, is a convolutional neural network which is GPU-friendly and can have shared CNN along with detection network.
• RPN introduces novel “anchor” boxes that serve as references at multiple scales and aspect ratios
• Propose alternating training scheme to unify RPN and Fast R-CNN object detection:
  • Alternating between fine-tuning for region proposals and object detection while keeping proposals fixed.