Advanced Deep Learning with Python

Detection network

Now that we've discussed the RPN, let's focus on the detection network. To do this, we'll go back to the diagram The structure of Faster R-CNN at the beginning of the Object detection with Faster R-CNN section. Let's recall that in the first stage, the RPN already generated the RoI coordinates. The detection network is a regular classifier, which determines the type of object (or background) in the current RoI. Both the RPN and the detection net share their first convolutional layers, borrowed from the backbone net. But the detection net also incorporates the proposed regions from the RPN, along with the feature maps of the last shared layer.

But how do we combine the inputs? We can do this with the help of Region of Interest ...