Where Global Context and Weight Sharing Work in Compact Vision Backbones1 versionW_d4nxyjwsv1 · current