As discussed in a previous Chapter 7, Image Segmentation, in the Deep instance segmentation recipe, region-based object detection methods (for example, R-CNN and Fast R-CNN) rely on region proposal algorithms (selective search) to guess object locations. Faster R-CNN is yet another region-based object detection model that was proposed as an improvement on R-CNN (2013) and Fast R-CNN (2015), by Girshick et al. again. Fast R-CNN decreases the execution time of detection (for example, for the slower R-CNN model) by introducing ROI Pooling, but still, region proposal computation becomes a bottleneck. Faster R-CNN introduces a Region Proposal Network (RPN). It achieves almost cost-free region proposals by sharing convolutional features with the detection network.
A Region Proposal Network (RPN) is an FCN that predicts regions that potentially contain...
