A NOVEL RANK CORRELATION MEASURE FOR MANIFOLD LEARNING ON IMAGE RETRIEVAL AND PERSON RE-ID

Abstract

Effectively measuring similarity among data samples represented as points in high-dimensional spaces remains a major challenge in retrieval, machine learning, and computer vision. In these scenarios, unsupervised manifold learning techniques grounded on rank information have been demonstrated to be a promising solution. However, various methods rely on rank correlation measures, which often depend on a proper definition of neighborhood size. On current approaches, this definition may lead to a reduction in the final desired effectiveness. In this work, a novel rank correlation measure robust to such variations is proposed for manifold learning approaches. The proposed measure is suitable for diverse scenarios and is validated on a Manifold Learning Algorithm based on Correlation Graph (CG). The experimental evaluation considered 6 datasets on general image retrieval and person Re-ID, achieving results superior to most state-of-the-art methods.