IDEAL: Independent domain embedding augmentation learning

Abstract

Deep metric learning is fundamental to open-set pattern recognition and has become a focal point of research in recent years. Significant efforts have been devoted to designing sampling, mining, and weighting strategies within algorithmic-level deep metric learning (DML) loss objectives. However, less attention has been paid to input-level but essential data transformations. In this paper, we develop a novel mechanism, independent domain embedding augmentation learning (IDEAL) method. It can simultaneously learn multiple independent embedding spaces for multiple domains generated by predefined data transformations. Our IDEAL is orthogonal to existing DML techniques and can be seamlessly combined with one DML approach for enhanced performance. Empirical results on visual retrieval tasks demonstrate the superiority of the proposed method. For instance, IDEAL significantly improves the performance of both Multi-Similarity (MS) Loss and Hypergraph-Induced Semantic Tuplet (HIST) loss. Specifically, it boosts the Recall$@1$ from 84.5% $\to$ 87.1% for MS Loss on Cars-196 and from 65.8% $\to$ 69.5% on CUB-200. Similarly, for HIST loss, IDEAL improves the performance on Cars-196 from 87.4% $\to$ 90.3%, on CUB-200 from 69.7% to 72.3%. It significantly outperforms methods using basic network architectures (e.g., ResNet-50, BN-Inception), such as XBM and Intra-Batch. The source code of our proposed method is available at https://github.com/emdata-ailab/Ideal-learning.