Parallelization strategies for markerless human motion capture

Markerless motion capture (MMOCAP) is the problem of determining the pose of a person from images captured by one or several cameras simultaneously without using markers on the subject. Evaluation of the solutions is frequently the most time-consuming task, making most of the proposed methods inapplicable in real-time scenarios. This paper presents an efficient approach to parallelize the evaluation of the solutions in CPUs and GPUs. Our proposal is experimentally compared on six sequences of the HumanEva-I dataset using the CMAES algorithm. Multiple algorithm’s configurations were tested to analyze the best trade-off with regard to the accuracy and computing time. The proposed methods obtain speedups of 8×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} in multi-core CPUs, 30×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} in a single GPU and up to 110×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} using 4 GPUs.