Dual Generative Models for Human Pose Estimation

Given a image photographed somebody in action, we describe a dual-generative-model approach for estimating human body pose from silhouette. In contrast to existing techniques, which mostly learn regression model whereby make inference of body pose for unknown input [1, 2, 6, 7], we transform the problem into searching of the best pair of upper pose and lower pose. This searching strategy can reduce significantly computational cost from global search of O(m̃n) to two subposes search of O(m + n) 1plus some cost of auxiliary works, however which can be extremely optimized. For regular cases, we observed that it is easy to divide image into two subimages representing upper and lower body. A local generative model is used to select from database enough candidates of upper and lower poses, and combine them under spatial constraint into full poses. Together with the full-body scores evaluated from a global generative model, we overall evaluate all full pose candidates and choose the best subpose pair accordingly. Examples are given of body pose estimation from both synthesized images and real images, both showing promising results.