Tripool: Graph triplet pooling for 3D skeleton-based action recognition

Highlights：

• Research highlights 2: The pooling method provides to optimize a graph triplet loss, in which both graph topology and graph context are captured by our pooling method. We also accelerate this process via minimizing its upper bound.

• Research highlights 3: We conduct extensive experiments to evaluate this method. Our network constructed with Tripool gets the state-of-the-art performance for the skeleton-based action recognition tasks on two current largest datasets.

摘要

•Research highlights 1: We propose a novel graph pooling method, and for the first time, introduce it into GCNs for skeleton-based action recognition. This method not only reduces the time complexity, as well increases the reception filed. Moreover, it breaks the structure constrain for the high-level features.•Research highlights 2: The pooling method provides to optimize a graph triplet loss, in which both graph topology and graph context are captured by our pooling method. We also accelerate this process via minimizing its upper bound.•Research highlights 3: We conduct extensive experiments to evaluate this method. Our network constructed with Tripool gets the state-of-the-art performance for the skeleton-based action recognition tasks on two current largest datasets.

论文评审过程：Received 8 July 2020, Revised 20 November 2020, Accepted 23 February 2021, Available online 6 March 2021, Version of Record 17 March 2021.