Metric Learning with Adaptive Density Discrimination

TL;DR

This paper introduces a novel metric learning method that models class distributions to improve local discrimination, achieving state-of-the-art results in fine-grained visual recognition and significantly reducing training time.

Contribution

It presents an adaptive density discrimination approach that explicitly models class distributions to enhance metric learning performance and efficiency.

Findings

Achieves state-of-the-art classification accuracy on fine-grained datasets.

Reduces training iterations by 5-30 times compared to triplet loss.

Improves attribute concentration and hierarchy recovery in learned representations.

Abstract

Distance metric learning (DML) approaches learn a transformation to a representation space where distance is in correspondence with a predefined notion of similarity. While such models offer a number of compelling benefits, it has been difficult for these to compete with modern classification algorithms in performance and even in feature extraction. In this work, we propose a novel approach explicitly designed to address a number of subtle yet important issues which have stymied earlier DML algorithms. It maintains an explicit model of the distributions of the different classes in representation space. It then employs this knowledge to adaptively assess similarity, and achieve local discrimination by penalizing class distribution overlap. We demonstrate the effectiveness of this idea on several tasks. Our approach achieves state-of-the-art classification results on a number of fine-grained visual recognition datasets, surpassing the standard softmax classifier and outperforming triplet loss by a relative margin of 30-40%. In terms of computational performance, it alleviates training inefficiencies in the traditional triplet loss, reaching the same error in 5-30 times fewer iterations. Beyond classification, we further validate the saliency of the learnt representations via their attribute concentration and hierarchy recovery properties, achieving 10-25% relative gains on the softmax classifier and 25-50% on triplet loss in these tasks.