Over time, Flickr has become a very large photo and video repository, claiming more than 13 billion photos available onsite. It sounds logical that many people are beginning to consider it as a treasure trove to explore human society. Real-world event discovery from Flickr, however, is inherently difficult because of the heterogeneity of its data. Methods to identify events from raw data already exist, unfortunately to little avail. This paper, thus, presents a three-stage framework to improve on these methods. In short, it is not a new method but a novel way to frame existing methods: Stage 1 consists of the design of a multimodal fusion model; Stage 2, the construction of a dual graph matrix factorization model; and Stage 3, event discovery via hybrid clustering algorithms. The paper starts with an extensive review of the state of the art, then describes the framework, and finally presents results from an extensive testing campaign.

The framework description starts from an interesting figure depicting its overall structure; from Flickr source data, multimodal by nature, features are extracted. Stage 1 consists of a voting mechanism generating a data graph and base graphs for these features. Stage 2 builds an event representation from those data and base graphs via a mechanism reinforcing previous voting results. This, by the way, is the major contribution of this paper to the field, according to its authors. Stage 3 groups different events in clusters. The paper describes all phases in mathematical terms, and a high-level algorithm is also given.

The testing part is also very detailed. First, it presents the equipment used, a dual six-core server and the dataset, publicly available data from a 2014 social event. The test consists of extracting images associated with this event from all images; it is performed first with algorithms already in use, which constitute the baseline results, and then the new framework is applied. In short, the tests demonstrate the framework to be very robust, meaning results do not vary much when choosing different parameters, which improve user friendliness, or number of images used, which reduces computational complexity and allows the use of small-sized datasets.

As with most research papers, this work builds on and extends previous results; what is interesting here is that these extensions seem to be easy to apply to existing methods, thus appealing not only to researchers who want to keep current on the state of the art in their field, but also to professionals in many fields, notably marketing departments, as a starting point to build their own analysis tools.