Recent progress in using recurrent neural networks (RNNs) for image description has motivated us to explore the application of RNNs to video description. Recent work has also suggested that attention mechanisms may be able to increase performance. To this end, we apply a long short-term memory (LSTM) network in two configurations: with a recently introduced soft-attention mechanism, and without. Our results suggest two things. First, incorporating a soft-attention mechanism into the text generation RNN significantly improves the quality of the descriptions. Second, using a combination of still frame features and dynamic motion-based features can also help. Ultimately, our combined approach exceeds the state-of-art on both BLEU and Meteor on the Youtube2Text dataset. We also present results on a new, larger and more complex dataset of paired video and natural language descriptions based on the use of Descriptive Video Service (DVS) annotations which are now widely available as an additional audio track on many DVDs.

这篇论文提出了一种基于循环神经网络和 3-D 卷积神经网络以及时间注意力机制的视频自动描述方法，并在Youtube2Text数据集上的BLEU和METEOR度量标准上超过了当前最先进的结果，并在更大更具挑战性的配对视频和自然语言描述数据集上进行了测试。

利用时间结构描述视频