Characterizing the impact of using features extracted from pre-trained models on the quality of video captioning sequence-to-sequence models.