On the automatic online collection of training data for visual event modeling

The last decade has witnessed the development and uprising of social media web services. The use of these shared online media as a source of huge amount of data for research purposes is still a challenging problem. In this paper, a novel framework is proposed to collect training samples from online media data to model the visual appearance of social events automatically. The visual training samples are collected through the analysis of the spatial and temporal context of media data and events. While collecting positive samples can be achieved easily thanks to dedicated event machine-tags, finding the most representative negative samples from the vast amount of irrelevant multimedia documents is a more challenging task. Here, we argue and demonstrate that the most common negative samples, originating from the same location as the event to be modeled, are best suited for the task. A novel ranking approach is devised to automatically select a set of negative samples. Finally the automatically collected samples are used to learn visual event models using Support Vector Machine (SVM). The resulting event models are effective to filter out irrelevant photos and perform with a high accuracy as demonstrated on various social events originating for various categories of events.