#feature-maps

Question 1: Single choice

A multiscale detector uses anchors on a large feature map near the input for small objects, and anchors on a small feature map near the output for large objects. Why do the deeper, smaller feature maps suit large objects?

Their units have smaller receptive fields, so they can concentrate on the fine details of large objects1
Their units carry fewer channels, so the predictors on top of them have fewer parameters to fit2
They are upsampled back to the input resolution before prediction, so they see more of the pixels3
Their units have larger receptive fields and more abstract features, so they can cover whole large objects4

Timed test