Lightweight channel-topology based adaptive graph convolutional network for skeleton-based action recognition
Kaixuan Wang, Hongmin Deng, Qilin Zhu
Abstract
With the development of graph convolutional network (GCN) over the recent years, skeleton-based action recognition has achieved satisfactory results. However, some existing GCN-based models were complex because of lots of parameters in the models. Moreover, a large proportion of the existing GCN-based extraction methods for temporal feature could not effectively extract temporal features. To address this problem, a lightweight channel-topology based adaptive graph convolutional network (LC-AGCN), is proposed in this paper. And it includes three innovative and important blocks. To be specific, firstly, the channel-topology adaptive graph convolution (CAGC) block is proposed for spatial feature extraction (SConv), and a modified multi-scale convolution block is introduced to extract temporal features (TConv). Then, in order to decrease the quantity of parameters, the bottleneck structure is introduced to lighten the model and obtain the desired result. Finally, in order to embody the principle of ”few parameters with high evaluating accuracy”, a parameter λap is creatively proposed to reflect the performance of lightweight models, which means the ratio of precision to parameter quantity. Extensive experiments demonstrate that our method greatly reduces the quantity of parameters of the model while ensuring high enough accuracy. The superiority of LC-AGCN has been proved on two large-scale public datasets named NTU-RGB+D and NTU-RGB+D 120, respectively.