Beoungwoo Kang

Attention is often redundant across feature scales in semantic segmentation, but this work shows you can propagate attention maps from the deepest layer to shallower layers and maintain accuracy while slashing compute.