Feed Two Birds with One Scone: Exploiting Wild Data for Both Out-of-Distribution Generalization and Detection