DensePoseGait: Dense Human Pose Part-Guided for Gait Recognition

Gait recognition is a technology that identifies human ID according to the human unique biometric gait feature. It has two popular categories, appearance-based and model-based algorithms. Appearance-based algorithms generally use human silhouettes as the initial input data. External factors such as clothing and physical carrying can drastically alter human silhouettes. In contrast, model-based algorithms tend to be more robust in regard to appearances, with human skeletons providing the initial input data in general. However, human skeletons suffer from limited information which causes an obstacle to increasing performance. In this paper, we, therefore, address this challenge by presenting two new databases, named CASIA-B-DensePose and MoBo-DensePose, which are based on the publicly available multiview database, CASIA-B and MoBo. They exploit UV coordinates of body surface and human semantic segmentation as the initial gait feature. It is less sensitive to human shape compared with human silhouettes, and has richer semantic information compared with human skeletons. In addition, we also introduce a novel model-based framework, DensePoseGait, to take full advantage of databases. Unlike traditional algorithms which either extract isolated local features or combine them with global features, DensePoseGait uses a novel way to exploit partial features. That is, human pose parts are employed as a regulator to guide the learning of global features in the training stage. Its core idea is to establish better representative features with the assistance of partial features, but not require additional calculation in the inference stage. We believe these databases and framework can offer researchers a fresh perspective on model-based gait recognition and inspire further exploration and advancements in this area.