View-invariant action recognition via Unsupervised AttentioN Transfer (UANT)

作者：

Highlights：

• We propose an Unsupervised AttentioN Transfer (UANT) network, that transfers attention from reference to arbitrary views to overcome view change.The UANT network learns a common attention map between different views via attention transfer. We exhaustively evaluate our approach on the UESTC and NTU dataset, and achieves outstanding recognition performance.Our proposed approach performs attention transfer in an unsupervised way, where view information is hidden.

摘要

•We propose an Unsupervised AttentioN Transfer (UANT) network, that transfers attention from reference to arbitrary views to overcome view change.The UANT network learns a common attention map between different views via attention transfer. We exhaustively evaluate our approach on the UESTC and NTU dataset, and achieves outstanding recognition performance.Our proposed approach performs attention transfer in an unsupervised way, where view information is hidden.

论文关键词：View-invariant recognition,Cross-view evaluation,Attention learning,Transfer learning

论文评审过程：Received 28 October 2019, Revised 15 December 2020, Accepted 26 December 2020, Available online 6 January 2021, Version of Record 29 January 2021.

论文官网地址：https://doi.org/10.1016/j.patcog.2020.107807