Abstract

In this work, we evaluate contrastive models for the task of imageretrieval. We hypothesise that models that are learned to encodesemantic similarity among instances via discriminative learningshould perform well on the task of image retrieval, where rele-vancy is defined in terms of instances of the same object. Throughour extensive evaluation, we find that representations from mod-els trained using contrastive methods perform on-par with (andoutperforms) a pre-trained supervised baseline trained on the Ima-geNet labels in retrieval tasks under various configurations. This isremarkable given that the contrastive models require no explicitsupervision. Thus, we conclude that these models can be used tobootstrap base models to build more robust image retrieval engines.