ROI Pooling

Region of interest pooling explained (deepsense.ai)

It's a NN layer taking two inputs:

• A fixed-size feature map from a CNN

• A N x 5 matrix of RoIs. With N = #RoI and 5 = #bbox coordinates + the image index

How does it work?

For every RoI, it takes the corresponding section of the feature map and scales it to a pre-defined size (eg 5x5). For it,

given a RoI proposed

divide the RoI into equal-sized sections

apply max-pooling on each section and append it to the output buffer

Note: RoI size does not need to be multiple of the grid size (eg a 7x5 region can be divided into a 2x2)