Deep Relational Metric Learning

TL;DR

This paper introduces a deep relational metric learning framework that adaptively models both interclass and intraclass variations by using an ensemble of features and relational inference, improving image clustering and retrieval.

Contribution

It proposes a novel ensemble-based relational module that captures correlations among features to enhance deep metric learning performance.

Findings

Improves clustering and retrieval accuracy on benchmark datasets.

Achieves state-of-the-art results compared to existing methods.

Effectively models intraclass variations to handle unseen classes.

Abstract

This paper presents a deep relational metric learning (DRML) framework for image clustering and retrieval. Most existing deep metric learning methods learn an embedding space with a general objective of increasing interclass distances and decreasing intraclass distances. However, the conventional losses of metric learning usually suppress intraclass variations which might be helpful to identify samples of unseen classes. To address this problem, we propose to adaptively learn an ensemble of features that characterizes an image from different aspects to model both interclass and intraclass distributions. We further employ a relational module to capture the correlations among each feature in the ensemble and construct a graph to represent an image. We then perform relational inference on the graph to integrate the ensemble and obtain a relation-aware embedding to measure the similarities. Extensive experiments on the widely-used CUB-200-2011, Cars196, and Stanford Online Products datasets demonstrate that our framework improves existing deep metric learning methods and achieves very competitive results.