Asymmetry quantification in cross modal retrieval using copulas

Abstract

Copulas are used to describe and explain the asymmetry between image-text and text-image retrieval observed in different values of the mean average precision (MAP). We use empirical copulas to quantify the asymmetry in a general framework of cross-modal retrieval via suitable asymmetry measures. Several experiments are done on real world dataset feautures to prove the relevance of our analysis.