Hyperbolic Image Embeddings

  • 2019-04-03 21:10:12
  • Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, Victor Lempitsky
  • 51

Abstract

Computer vision tasks such as image classification, image retrieval andfew-shot learning are currently dominated by Euclidean and sphericalembeddings, so that the final decisions about class belongings or the degree ofsimilarity are made using linear hyperplanes, Euclidean distances, or sphericalgeodesic distances (cosine similarity). In this work, we demonstrate that inmany practical scenarios hyperbolic embeddings provide a better alternative.

 

Quick Read (beta)

loading the full paper ...