The assignment of importance scores to particular frames or (short) segments\nin a video is crucial for summarization, but also a difficult task. Previous\nwork utilizes only one source of visual features. In this paper, we suggest a\nnovel model architecture that combines three feature sets for visual content\nand motion to predict importance scores. The proposed architecture utilizes an\nattention mechanism before fusing motion features and features representing the\n(static) visual content, i.e., derived from an image classification model.\nComprehensive experimental evaluations are reported for two well-known\ndatasets, SumMe and TVSum. In this context, we identify methodological issues\non how previous work used these benchmark datasets, and present a fair\nevaluation scheme with appropriate data splits that can be used in future work.\nWhen using static and motion features with parallel attention mechanism, we\nimprove state-of-the-art results for SumMe, while being on par with the state\nof the art for the other dataset.\n