Reinforcement Learning with TensorFlow

Faster R-CNN

Faster R-CNN goes by its name. It's faster than previous Fast R-CNNs. This was achieved by replacing the slowest part of Fast R-CNN, selective search for generation box regions (of interest) with a very small convolution network called Regional Proposal Network to perform the same task, generating box regions that are highly probable to contain the object (regions of interest).

Faster R-CNN implements the idea of anchor boxes to handle the variations in aspect ratio and scale of objects. For each region there are three anchor boxes for scale and three aspect ratios. Therefore, for each location we have nine boxes fed to the Regional Proposal Network (RPN) predicting the probability of the region being a background or a foreground. ...