Overview

Metric learning studies how to shape similarity so that distances in a representation reflect the relationships that matter for a task. Common approaches include contrastive, triplet, proxy, and angular-margin losses, alongside learned Mahalanobis, manifold, and optimal-transport distances. These methods support image and cross-modal retrieval, person re-identification, few-shot classification, clustering, and open-set recognition, with evaluations often measuring retrieval ranking, recognition accuracy, and robustness to domain shift. A central challenge is making learned geometries expressive enough to capture task-specific structure while remaining stable under scarce or noisy labels and generalizing to unfamiliar data.

Latest papers

CardsList

92 papers11 papers in the last four weeks, up 10% on the four weeks before. 0.1% of all new papers.

  1. Why and How: Knowledge-Guided Learning for Cross-Spectral Image Patch Matching

    Dec 15, 2024Chuang Yu, Yunpeng Liu, Jinmiao Zhao +5Image MatchingCross-Modal Feature Matching

  2. Metric Learning with Quantum Feature Spaces

    Dec 4, 2023Vinayak Sharma, Lord Sen, Ashish Padhy +4Quantum Machine LearningMetric Learning