Coarse localization using space-time and semantic-context representations of geo-referenced video sequences

We introduce a new video contents description approach and use it for the purpose of coarse localization. It is based on a Bag of Words representation combining both space-time STIP features and semantic-context SSC features. We assume that adding semantic context encodes in a more efficient way the spatio-temporal information into video sequences. The resulting augmented descriptor is related to a geographic location that can be estimated within a classic classification framework. We show that on real geo-referenced video sequences, the proposed system improves the localization compared to classical descriptors.