#feature-pyramid-network

Question 1: Single choice

A Feature Pyramid Network (FPN) adds a top-down pathway with lateral connections to a ResNet backbone and predicts from the resulting maps P2–P5 instead of the backbone's own maps C2–C5. What problem does this solve?

Fine maps such as C2 have weak semantics; the top-down path adds deep semantics to them1
The backbone maps are too small for anchors, so FPN upsamples every level to the input resolution2
It saves computation by skipping the deepest backbone stage C53
It removes the need for anchors, since each pyramid level predicts boxes directly4

#feature-pyramid-network

Timed test