Key points are not available for this paper at this time.
Many automatic image annotation methods are based on the learning by example paradigm. Image tagging, through manual image inspection, is the first step towards this end. However, manual image annotation, even for creating the training sets, is time-consuming, complicated and contains human subjectivity errors. Thus, alternative ways for automatically creating training examples, i.e., pairs of images and tags, are crucial. As we showed in one of our previous studies, tags accompanying photos in social media and especially the Instagram hashtags can be used for image annotation. However, it turned out that only a 20% of the Instagram hashtags are actually relevant to the content of the image they accompany. Identifying those hashtags through crowdsourcing is a plausible solution. In this work, we investigate the effectiveness of the HITS algorithm for identifying the right tags in a crowdsourced image tagging scenario. For this purpose, we create a bipartite graph in which the first type of nodes corresponds to the annotators and the second type to the tags they select, among the hashtags, to annotate a particular Instagram image. From the results, we conclude that the authority value of the HITS algorithm provides an accurate estimation of the appropriateness of each Instagram hashtag to be used as a tag for the image it accompanies while the hub value can be used to filter out the dishonest annotators.
Giannoulakis et al. (Wed,) studied this question.