Inferring generic activities and events from image content and bags of geo-tags

The use of contextual information in building concept detectors for digital media has caught the attention of the multimedia community in the recent years. Generally speaking, any information extracted from image headers or tags, or from large collections of related images and used at classification time, can be considered as contextual. Such information, being discriminative in its own right, when combined with pure content-based detection systems using pixel information, can improve the overall recognition performance significantly. In this paper, we describe a framework for probabilistically modeling geographical information using a Geographical Information Systems (GIS) database for event and activity recognition in general-purpose consumer images, such as those obtained from Flickr. The proposed framework discriminatively models the statistical saliency of geo-tags in describing an activity or event. Our work leverages the inherent patterns of association between events and their geographical venues. We use descriptions of small local neighborhoods to form bags of geo tags as our representation. Statistical coherence is observed in such descriptions across a wide range of event classes and across many different users. In order to test our approach, we identify certain classes of activities and events wherein people commonly participate and take pictures. Images and corresponding metadata, for the identified events and activities, are obtained from Flickr. We employ visual detectors obtained from Columbia University (Columbia 374), which perform pure visual event and activity recognition. In our experiments, we present the performance advantage obtained by combining contextual GPS information with pixel-based detection systems.