Weakly-supervised action localization via embedding-modeling iterative optimization

作者：

Highlights：

• An iterative optimizing mechanism is designed for effective action recognition with trimmed-untrimmed parallel modeling.

• A shared subspace embedding is proposed with generative adversarial networks for coherent knowledge transfer.

• A two-stage self-attentive representation learning workflow is developed to capture the fine-grained frame-level relevance.

• Extensive experiments are conducted on two challenging untrimmed video datasets with superior results.

摘要

•An iterative optimizing mechanism is designed for effective action recognition with trimmed-untrimmed parallel modeling.•A shared subspace embedding is proposed with generative adversarial networks for coherent knowledge transfer.•A two-stage self-attentive representation learning workflow is developed to capture the fine-grained frame-level relevance.•Extensive experiments are conducted on two challenging untrimmed video datasets with superior results.

论文关键词：Action recognition,Temporal action localization,Attention mechanism,Generative adversarial networks,Subspace embedding

论文评审过程：Received 9 October 2019, Revised 24 September 2020, Accepted 14 December 2020, Available online 16 January 2021, Version of Record 21 January 2021.

论文官网地址：https://doi.org/10.1016/j.patcog.2021.107831