Augmenting training sets with still images for video concept detection

Accessing the visual information of video content is a challenging task. Automatic annotation techniques have made significant progress, however they still suffer from the lack of appropriate training data. To overcome this problem we propose the use of still images taken from a photo sharing website as an additional resource for training. However, a mere extension of the training set with still images does not yield a large gain in classification accuracy. We show that using a combination of techniques for bridging the differences between still images and video keyframes improves classification performance compared to simply augmenting the training set.