Abstract
This paper proposes dynamically modulated metric learning (DMML) for learning a tiered similarity space to perform visual search. Existing methods often treat the training samples having different degree of information with equal importance which hinders in capturing the underlying granularities in visual similarity. Proposed DMML automatically exploits the informativeness of samples during training by leveraging correlation between image attributes and embedding that are learned jointly. The two tasks are interlinked by supervising signals where the predicted attribute vectors are used to dynamically learn the loss function. To this end, we propose a new soft-binomial deviance loss that helps to capture the feature similarity space at multiple granularities. Compared to recent ensemble and attention based methods, our DMML framework is conceptually simple yet effective, and achieves state-of-the-art performances on standard benchmark datasets; e.g. an improvement of 4% Recall@1 over the SOTA [1] on DeepFashion dataset.