A Dataset of Image Pairs Related by Homographies for Use in Image Matching

Few resources exist for the purpose of testing algorithms for feature point matching, since most image pairs don't easily admit an automatic way of calculating the accuracy of a proposed match between them. The Murals dataset contains eight image pairs gathered from Flickr. The images from each pair are related by a homography resulting from perspective change. The selected images exhibit repetitive features to further enhance the difficulty of matching.


The images are provided with a set of ground truth homographies. To increase the amount of possible test cases, we also provide a python script, which can create a set of cropped pairs with different degrees of overlap and calculate a homography constituting the actual transformation between each pair. This will approximate testing on a larger set of similar images with partially overlapping pairs (the paper below explains this in more detail).

The dataset is released under a creative commons license and can be downloaded here: The zip-file is encrypted; please email us for the password.

More details about the dataset can be found in the following paper:

J. T. Arnfred, S. Winkler, S. Süsstrunk.
Mirror Match: Reliable feature point matching without geometric constraints.
Proc. 2nd IAPR Asian Conference on Pattern Recognition (ACPR), Okinawa, Japan, Nov. 5-8, 2013.

Please cite the above paper if you use the Murals dataset.