Datasets for Asymmetric Image Comparisons

We present two datasets to evaluate asymmetric image comparisons, in the context of retrieval applications:
(i) Asym-QCD: this dataset considers the case where the query image is contained in a database image
(ii) Asym-DCQ: this dataset considers the case where a database image is contained in the query image

The images used in this dataset are collected from the Stanford Mobile Visual Search, INRIA Holidays and MIR-FLICKR datasets.
Instructions are given (see README.txt file) for appropriate download and setup of dataset images.

The query images used in Asym-QCD are clean images of objects, and their corresponding correct database matches are images where the object is shown along with clutter.
For the Asym-DCQ dataset, these two sets of images are reversed: the database contains clean object images, and the queries are cluttered images.
We construct several versions of each of the two datasets, with different amounts of clutter:
(i) Asym-QCD: from 0 to 40 clutter images are added to each database item,
(ii) Asym-DCQ: from 0 to 40 clutter images are added to each query item.