The Hubble Image Similarity Project

Abstract We have created a large database of similarity information between subregions of Hubble Space Telescope images. These data can be used to assess the accuracy of image-search algorithms based on computer vision methods. The images were compared by humans in a citizen science project, where they were asked to select similar images from a comparison sample. We utilized the Amazon Mechanical Turk system to pay our reviewers a fair wage for their work. Nearly 850,000 comparison measurements have been analyzed to construct a similarity distance matrix between all the pairs of images. We describe the algorithm used to extract a robust distance matrix from the (sometimes noisy) user reviews. The results are very impressive: The data capture similarity between images based on morphology, texture, and other details that are sometimes difficult even to describe in words (e.g., dusty absorption bands with sharp edges). The collective visual wisdom of our citizen scientists matches the accuracy of the trained eye, with even subtle differences among images faithfully reflected in the distances.