Unsupervised image embedding using nonparametric statistics

Abstract

Embedding images into a low dimensional space has a wide range of applications: visualization, clustering, and pre-processing for supervised learning. Traditional dimension reduction algorithms assume that the examples densely populate the manifold. Image databases tend to break this assumption, having isolated islands of similar images instead. In this work, we propose a novel approach that embeds images into a low dimensional Euclidean space, while preserving local image similarities based on their scale invariant feature transform (SIFT) vectors. We make no neighborhood assumptions in our embedding. Our algorithm can also embed the images in a discrete grid, useful for many visualization tasks. We demonstrate the algorithm on images with known categories and compare our accuracy favorably to those of competing algorithms.