Video summarization with u-shaped transformer

In recent years, supervised video summarization has made tremendous progress with treating it as a sequence-to-sequence learning task. However, traditional recurrent neural networks (RNNs) have limitations in sequence modeling of long sequences, and the use of a transformer for sequence modeling requires a large number of parameters. We propose an efficient U-shaped transformer for video summarization tasks in this paper to address this issue, which we call "Uformer". Precisely, Uformer consists of three key components: embedding, Uformer block, and prediction head. First of all, the image features sequence is represented by the pre-trained deep convolutional network, then represented by a liner embedding. The image feature sequence differences are also represented by another liner embedding and concatenate together to form a two-stream embedding feature in the embedding component. Secondly, we stack multiple transformer layers into a U-shaped block to integrate the representations learned from the previous layers. Multi-scale Uformer can not only learn longer sequence information but also reduce the number of parameters and calculations. Finally, prediction head regression the localization of the keyframes and learning the corresponding classification scores. Uformer combine with non-maximum suppression (NMS) for post-processing to get the final video summarization. We improved the F-score from 50.2% to 53.9% by 3.7% on the SumMe dataset and improved F-score from 62.1% to 63.0% by 0.9% on the TVSum dataset. Our proposed model with 0.85M parameters which are only 32.32% of DR-DSN's parameters.