Abstract

Gait pattern is one of the most promising biometrics for practical applications since it can be captured at a long distance without requiring intentional cooperation. It is a feasible solution for human identification in transformer substations of power grid. In the latest literature, the basic components of most appearance-based methods can be straightly concluded into three categories, i) the Backbone to extract gait feature maps, ii) the Neck for spatio-temporal feature aggregation, iii) the Head for metric space mapping. Notably, previous works paid much attention to the Backbone and Neck but ignored the significant Head designs. The majority of works separated holistic representations into multiple partial features, but existing Heads are treating all parts equally. However, we argue that treating all parts equivalent in Head seems counterintuitive since it is natural for humans to distinguish two individuals by considering the most significant patterns rather than paying equal attention to each part. To address these issues, we proposed a universal Head named GaitHead, which can be easily applied in most part-based gait recognition networks to adaptively adjust the cross-samples attentive scores among identifying samples according to the identity, camera viewpoint, and walking condition. Specifically, we make the efforts in two aspects for the GaitHead. On the one hand, the Cross-Part Attention Encoder (CPAE) module integrates global information into the part-level attentive representations to obtain the characteristics of the identity, camera viewpoint, and walking condition for each part. On the other hand, the Part Aware Triplet Loss (PAT Loss) is proposed to supervise the cross-sample attentive scores in a well-designed way. Experiments on two common public databases, CASIA-B and OUMVLP, have fully proved that GaitHead is a plug-and-play Head for most part-based gait recognition networks to improve their performance considerably.

Statistics