#region-proposal-network

Faster R-CNN's region proposal network (RPN) places $k = 9$ anchors (3 scales × 3 aspect ratios) at every position of a $40 \times 60$ convolutional feature map, so there are anchors in total. On top of the 512-dimensional features of VGG-16, two sibling $1 \times 1$ convolutions output $2k$ objectness scores and $4k$ box coordinates per position; together they have weights (ignoring biases).