#roialign

Question 1: Single choice

Mask R-CNN extracts the features of each region of interest (RoI) from a feature map with stride 16. RoIPool, used by Faster R-CNN, would map an RoI's left edge at $x = 37.8$ input pixels to $[37.8 / 16] = [2.36] = 2$ on the feature map, where $[\cdot]$ is rounding. What does Mask R-CNN's RoIAlign do instead?

Uses $2.36$ unrounded and reads the features there by bilinear interpolation1
Rounds the same way, but splits each RoI into more bins so the error averages out2
Resizes the image so that every RoI edge falls on a whole feature-map cell3
Rounds the same way, but averages each bin instead of taking its maximum4

Timed test