Both R-CNN and Fast R-CNN rely on a selective search method to develop a 2,000 region proposal, which results in a detection rate of 2 seconds per image compared to 0.2 seconds per image for most efficient detection methods. Shaoquing Ren, Kaiming He, Ross Girshick, and Jian Sun wrote a paper titled Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks to Improve the R-CNN Speed and Accuracy for Object Detection. You can read the paper at https://arxiv.org/abs/1506.01497.
The following diagram shows the architecture of faster R-CNN:
The key concepts are shown in the following list:
- Introduction of the input image to a Region Proposal Network (RPN), which outputs a set of rectangular region proposals for a given image.
- The RPN shares convolutional layers with state-of-the-art object detection networks.
- The RPN is trained by back...
