Multi-View Large Population Gait Database With Human Meshes and Its Performance Evaluation

Existing model-based gait databases provide the 2D poses (i.e., joint locations) extracted by general pose estimators as the human model. However, these 2D poses suffer from information loss and are of relatively low quality. In this paper, we consider a more informative 3D human mesh model with parametric pose and shape features, and propose a multi-view training framework for accurate mesh estimation. Unlike existing methods, which estimate a mesh from a single view and suffer from the ill-posed estimation problem in 3D space, the proposed framework takes asynchronous multi-view gait sequences as input and uses both multi-view and single-view streams to learn consistent and accurate mesh models for both multi-view and single-view sequences. After applying the proposed framework to the existing OU-MVLP database, we establish a large-scale gait database with human meshes (i.e., OUMVLP-Mesh), containing over 10,000 subjects and up to 14 view angles. Experimental results show that the proposed framework estimates human mesh models more accurately than similar methods, providing models of sufficient quality to improve the recognition performance of a baseline model-based gait recognition approach.