Multi Proxy Anchor Loss and Effectiveness of Deep Metric Learning Performance Metrics

Deep metric learning (DML) learns the mapping, which maps into embedding space in which similar data is near and dissimilar data is far. However, conventional proxy-based losses for DML have two problems: gradient problems and applying the real-world dataset with multiple local centers. Besides, DML performance metrics also have some issues have stability and flexibility. This paper proposes multi-proxies anchor (MPA) loss and normalized discounted cumulative gain (nDCG@k) metric. This study contributes three following: (1) MPA loss is able to learn the real-world dataset with multi-local centers. (2) MPA loss improves the training capacity of a neural network owing to solving the gradient issues. (3) nDCG@k metric encourages complete evaluation for various datasets. Finally, we demonstrate MPA loss's effectiveness, and MPA loss achieves higher accuracy on two datasets for fine-grained images.