GaitMG: A Multi-grained Feature Aggregate Network for Gait Recognition

In order to make better use of the temporal information of gait data and improve the accuracy of gait recognition. In this paper, we propose a multi-grained feature aggregate network(GaitMG), which contains two important modules: Multi-Grained Feature Aggregator(MGFA) and Spatio-Temporal Feature Fusion Module(STFFM). The MGFA: a novel applying of convolution, can tackle the problem of poor representation ability of single grained temporal features. STFFM can fuse the multi-grained temporal features obtained by MGFA, get a discriminative representation. On CASIA-B, our method can achieve rank-1 accuracy of 98.0% under normal walking condition, 94.1% under bag-carrying condition, 85.3% under coat-wearing condition.