Rethinking preventing class-collapsing in metric learning with margin-based losses

Elad Levi, Tete Xiao, Xiaolong Wang, Trevor Darrell