Deep Metric Learning with BIER: Boosting Independent Embeddings Robustly

摘要

Learning similarity functions between image pairs with deep neural networksyields highly correlated activations of embeddings. In this work, we show howto improve the robustness of such embeddings by exploiting the independencewithin ensembles. To this end, we divide the last embedding layer of a deepnetwork into an embedding ensemble and formulate training this ensemble as anonline gradient boosting problem. Each learner receives a reweighted trainingsample from the previous learners. Further, we propose two loss functions whichincrease the diversity in our ensemble. These loss functions can be appliedeither for weight initialization or during training. Together, ourcontributions leverage large embedding sizes more effectively by significantlyreducing correlation of the embedding and consequently increase retrievalaccuracy of the embedding. Our method works with any differentiable lossfunction and does not introduce any additional parameters during test time. Weevaluate our metric learning method on image retrieval tasks and show that itimproves over state-of-the-art methods on the CUB 200-2011, Cars-196, StanfordOnline Products, In-Shop Clothes Retrieval and VehicleID datasets.