Semantic Segmentation by Early Region Proxy

Yifan Zhang, Bo Pang, Cewu Lu

Typical vision backbones manipulate structured features. As a compromise, semantic segmentation has long been modeled as per-point prediction on dense regular grids. In this work,…