#skip-connections

Question 1: Single choice

FCN-32s upsamples its stride-32 class scores straight back to the image size, and its masks are coarse. How do FCN-16s and FCN-8s produce finer masks?

They add class scores from pool4 and pool3 (strides 16 and 8) to the upsampled coarse scores1
They are trained on larger input images, so that the stride-32 map contains more pixels2
They remove all pooling from the backbone so that every layer keeps the full input resolution3
They replace the final upsampling by a fully connected layer that outputs one score per pixel4

Timed test