How do people perceive image similarity?
Krzysztof Michalak, Bartlomiej Dzienkowski, Elzbieta Hudyma
(Wroclaw University of Technology, Poland)
Abstract:
In the presentation a problem of inter-rater agreement is discussed in
the case of human observers who judge how similar pairs of images are. In
our research we frequently refer to the concept of image similarity. To
better understand what images are perceived by people as similar, we
performed a series of tests in which participants were asked to assign a
level of similarity (from a predefined set of grades) to pairs of images
displayed on the screen. Image sets used in the experiments were
diversified with respect to the topic and with respect to the presence of
easily distinguishable objects. The results were analyzed using
inter-rater agreement measures such as Kappa coefficients. We discovered
that the perceived level of image similarity varies greatly among users.
We have even came across a pair of images to which each of the grades
(ranging from "not similar" to "very similar") was assigned by various
people exactly the same number of times. We also observed that the level
of agreement in the group of people depends on the characteristics of
images in a given image set, especially on the presence of easily
distinguishable objects. In the presentation the above issues will be
discussed. We will present the system developed at out Institute which
allows to perform the tests and to collect and analyze the results.
Collected data and processing methods will be presented and we will
discuss the results and conclusions