Project topics and materials
Get Complete Project Material File(s) Now! »
Topic Detection and Tracking
The Topic Detection and Tracking (TDT) initiative was first intended to explore techniques for identifying events and tracking their reappearance and evolution in a text document stream. Within the TDT context, an event was initially defined as “some unique thing that happens at some point in time” [3]. This definition was further extended to include location as well [104], defining an event as “something that happens at some specific time and place”. Under this definition, the World Trade Center attacks that took place on September 11, 2001 is an event. However, the media also reported the subsequent collapse of the World Trade Center towers. Here, it is unclear whether the events should be considered as separate events or whether they form part of one single event.
To address such an ambiguity, an amended definition was proposed in [2] stating that an event is “a specific thing that happens at a specific time and place along with all necessary preconditions and unavoidable consequences”. Although this definition makes some clarifications regarding event boundaries, it does not cover all possible types of event, since some of the necessary preconditions and unavoidable consequences may be ambiguous, unknown or subject to debate. Although the TDT-inspired definitions of an event introduce some useful concepts, they do not cover all possible types of events.
Towards event centric content organization in social media
Multimedia documents in User Generated Content (UGC) websites, as well as in personal collections, are often organized into events. Users are usually more likely to upload or gather pictures related to the same event, such as a given holiday trip, a music concert, a wedding, etc. This also applies to professional contents such as journalism or historical data that are even more systematically organized according to hierarchies of events.
Given a query event record represented by a set of photos, our method aims to retrieve other records of the same event, notably those generated by other actors or witnesses of the same real-world event. An illustration of two matching event records is presented in Figure 3.1. It shows how a small subset of visually similar and temporally coherent pictures might be used to match the two records, even if they include other distinct pictures covering different aspects of the event. Application scenarios related to such a retrieval paradigm are numerous. By simply uploading their own record of an event users might, for example, gain access to the community of other participants. They can then revive the event by browsing or collecting new data complementary to their own view of the event. If some previous event’s records had already been uploaded and annotated, the system might also automatically annotate a new record or suggest some relevant tags. The proposed method might also have nice applications in the context of citizen journalism. Automatically detecting the fact that a large number of amateur users did indeed record data about the same event would be very helpful for professional journalists in order to cover breaking news. Finally, tracking events across different media has a big potential for historians, sociologists, politicians, etc.
Visual based Event Matching
We first describe the proposed method in the general context of event records composed of a set of geo-tagged and time coded pictures. We further restrict ourselves to time coded only pictures since our experimental dataset did not include geo-tags.
Table of contents :
Résumé
1 Motivations
2 Problématiques
3 Contributions
3.1 Événement et instances d’événement
3.2 Recherche d’événements par similarité visuelle
3.3 Construction scalable et distribuée du graphe de similarité visuelle
3.4 Sélection de contenu
1 General Introduction
2 Events in Social Media
1 Events in the literature
1.1 Topic Detection and Tracking
1.2 Event Extraction
1.3 Multimedia Event Detection
1.4 Social Event Detection
2 Events in social media
3 Related tasks
3.1 Event matching
3.2 Content Selection
4 Conclusion
3 Event Identification in Social Media
1 Towards event centric content organization in social media
2 Visual based Event Matching
3 Enabling scalability
3.1 Multi-Probe LSH
3.2 The MapReduce framework
3.3 Multi-Probe LSH in the MapReduce framework
4 Experiments
4.1 Experimental settings
4.2 Results
4.3 Dicussion
5 Conclusion
4 Distributed k-NN Graphs construction
1 Problem Statement
2 Hashing-based K-NNG construction
2.1 Notations
2.2 LSH based K-NNG approximation
2.3 Balancing issues of LSH-based K-NNG
3 Proposed method
3.1 Random Maximum Margin Hashing
3.2 RMMH-based K-NNG approximation
3.3 Split local joins
3.4 MapReduce Implementation
4 Experimental setup
4.1 Datasets & Baselines
4.2 Performance measures
4.3 System environment
5 Experimental results
5.1 Hash functions evaluation
5.2 Experiments in centralized settings
5.3 Performance evaluation in distributed settings
6 Conclusion
5 Content Suggestion and Summarization
1 Content suggestion and summarization in UGC
1.1 Content Selection
1.2 Event Summarization
1.3 Content Suggestion
2 Building the Records Graph
3 Experiments
3.1 Experimental setup
3.2 Results
4 Conclusion
6 Related Work
1 Event Identification in Social Media
2 Event summarization
3 Large-scale k-NN Graph construction
4 Nearest Neighbors search
4.1 Curse of dimensionality
4.2 Approximate similarity search
Bibliography
