Multi-level Distance Regularization for Deep Metric Learning

We propose a novel distance-based regularization method for deep metric learning called Multi-level Distance Regularization (MDR). MDR explicitly disturbs procedure by regularizing pairwise distances between embedding vectors into multiple levels that represents degree of similarity pair. In the training stage, model is trained with both and an existing loss function learning, simultaneously; two losses interfere objective each other, it makes process difficult. Moreover, prevents some examples from being ignored or overly influenced in process. These allow parameters network to be settle on local optima better generalization. Without bells whistles, simple Triplet achieves the-state-of-the-art performance various benchmark datasets: CUB-200-2011, Cars-196, Stanford Online Products, In-Shop Clothes Retrieval. extensively perform ablation studies its behaviors show effectiveness MDR. By easily adopting our MDR, previous approaches can improved generalization ability.