Knowledge transfer from large teacher models to smaller student models has\nrecently been studied for metric learning, focusing on fine-grained\nclassification. In this work, focusing on instance-level image retrieval, we\nstudy an asymmetric testing task, where the database is represented by the\nteacher and queries by the student. Inspired by this task, we introduce\nasymmetric metric learning, a novel paradigm of using asymmetric\nrepresentations at training. This acts as a simple combination of knowledge\ntransfer with the original metric learning task.\n We systematically evaluate different teacher and student models, metric\nlearning and knowledge transfer loss functions on the new asymmetric testing as\nwell as the standard symmetric testing task, where database and queries are\nrepresented by the same model. We find that plain regression is surprisingly\neffective compared to more complex knowledge transfer mechanisms, working best\nin asymmetric testing. Interestingly, our asymmetric metric learning approach\nworks best in symmetric testing, allowing the student to even outperform the\nteacher.\n