Model Cascades for Efficient Image Search

03/27/2023

∙

by   Robert Hönig, et al.

∙

0

∙

share

Modern neural encoders offer unprecedented text-image retrieval (TIR) accuracy. However, their high computational cost impedes an adoption to large-scale image searches. We propose a novel image ranking algorithm that uses a cascade of increasingly powerful neural encoders to progressively filter images by how well they match a given text. Our algorithm reduces lifetime TIR costs by over 3x.