Hardness-Aware Deep Metric Learning

This paper presents a hardness-aware deep metric learning (HDML) framework for image clustering and retrieval. Most existing methods employ the hard negative mining strategy to alleviate lack of informative samples training. However, this only utilizes subset training data, which may not be enough characterize global geometry embedding space comprehensively. To address problem, we perform linear interpolation on embeddings adaptively manipulate their hardness levels generate corresponding label-preserving synthetics recycled so that information buried in all can fully exploited is always challenged with proper difficulty. As single synthetic each sample still describe unobserved distributions data crucial generalization performance, further extend HDML multiple sample. We propose randomly (HDML-R) method an (HDML-A) random adaptive directions, respectively, synthesis. Since generated might useful adaptive, selection three criteria qualified are beneficial metric. Extensive experimental results widely used CUB-200-2011, Cars196, Stanford Online Products, In-Shop Clothes Retrieval, VehicleID datasets demonstrate effectiveness proposed framework.