Learning Embeddings for Image Clustering: An Empirical Study of Triplet Loss Approaches

TL;DR

This paper empirically compares different Triplet Loss formulations for image embedding learning, evaluating their effectiveness in clustering tasks like k-means and correlation clustering, and introduces a new, improved Triplet Loss variant.

Contribution

It introduces a novel Triplet Loss formulation that outperforms existing methods in clustering quality and provides a comprehensive empirical analysis of Triplet Loss approaches for image clustering.

Findings

The new Triplet Loss formulation improves clustering performance.

Triplet Loss embeddings enhance discriminative feature learning.

Empirical results on CIFAR-10 demonstrate the effectiveness of the proposed method.

Abstract

In this work, we evaluate two different image clustering objectives, k-means clustering and correlation clustering, in the context of Triplet Loss induced feature space embeddings. Specifically, we train a convolutional neural network to learn discriminative features by optimizing two popular versions of the Triplet Loss in order to study their clustering properties under the assumption of noisy labels. Additionally, we propose a new, simple Triplet Loss formulation, which shows desirable properties with respect to formal clustering objectives and outperforms the existing methods. We evaluate all three Triplet loss formulations for K-means and correlation clustering on the CIFAR-10 image classification dataset.