International Conference on Computer and Knowledge Engineering

Home / 14th International Conference on Computer and Knowledge Engineering

Cluster Sampling: A Cluster-Driven Sampling Strategy for Deep Metric Learning

Authors :

Hamideh Rafiee¹ Ahmad Ali Abin² Seyed Soroush Majd³

1- Computer Science and Engineering Department, Shahid Beheshti University, Tehran, Iran 2- Computer Science and Engineering Department, Shahid Beheshti University, Tehran, Iran 3- Computer Science and Engineering Department Shahid Beheshti University

Keywords :

Deep Learning،Metric Learning،Deep Metric Learning،Sampling،Clustering

Abstract :

Abstract—Deep metric learning has gained significant attention recently due to its promising performance in image retrieval, face recognition, and clustering tasks. Deep metric learning algorithms map the original data from the initial feature space to a new embedded space by learning a mapping function, where the discriminative power of the samples is increased. Sampling strategies are a critical aspect of deep metric learning algorithms, as they determine the choice of training samples that shape the network’s learning process. This research introduces a novel sampling approach designed to optimize this selection for improved model performance. Given the high dimensionality of embedded spaces, choosing a sampling strategy is critical for enhancing the quality of learned representations and improving model performance. However, existing sampling methods often face significant challenges, such as slow convergence and the tendency to get trapped in suboptimal local minima. To address these challenges, this research proposes a novel sampling approach that leverages the clustering of sample representations within training batches to identify and select the most informative samples for effective training. This approach ensures a more dynamic learning process by clustering data based on shared features and selecting samples both within clusters and between related clusters. The selected samples enable the model to better learn the underlying data structure, thereby enhancing the discriminative power and robustness of the learned metrics. The effectiveness of the proposed approach is validated through empirical evaluations on the cars196 dataset, where it outperforms existing sampling methods.