Specific video identification via joint learning of latent semantic concept, scene and temporal structure.