Hierarchical Multi-scale Attention Networks for action recognition

作者：

Highlights：

•

• A Hierarchical Multi-Scale Attention Network (HM-AN) is proposed.

• Four methods of realizing attention mechanism for action recognition in video is proposed.

• By introducing Gumbel-softmax into HM-RNN, we make the stochastic neurons trainable by back propagation.

• Through adaptive temperature, we obtain improved results using Gumbel-softmax.

• Through visualizations, we provide insights for further research.

摘要

•A Hierarchical Multi-Scale Attention Network (HM-AN) is proposed.•Four methods of realizing attention mechanism for action recognition in video is proposed.•By introducing Gumbel-softmax into HM-RNN, we make the stochastic neurons trainable by back propagation.•Through adaptive temperature, we obtain improved results using Gumbel-softmax.•Through visualizations, we provide insights for further research.

论文关键词：Action recognition,Hierarchical multi-scale RNNs,Attention mechanism,Stochastic neurons

论文评审过程：Received 12 August 2017, Revised 21 November 2017, Accepted 22 November 2017, Available online 29 November 2017, Version of Record 1 December 2017.

论文官网地址：https://doi.org/10.1016/j.image.2017.11.005