Efficient Scale-Permuted Backbone with Learned Resource Distribution