2D Human Pose Estimation from Monocular Images: A Survey

Abstract

Human pose estimation is a computer vision problem that tries to estimate human body joints location and decide how they are connected to each other. It has been long studied and is still a frontier research field nowadays. In this paper, a comprehensive survey of human pose estimation from 2D monocular images is given, including from the classical representative works to the most recent deep-learning-based methods. The goal of the paper is to let the reader get a brief understanding on how human pose estimation methods work and see how these methods have developed, how they different from each other but in the same time share some common ideas. This paper inherits one of the most admitted way to categorize human pose estimation methods by dividing them into top-down and bottom-up methods, pipelines and some of the milestone works are introduced, comparison and discussion among ideas and methods are made. There are also new methods that jump out of the restriction of purely top- down or bottom-up, this paper includes them as well in later sections. Then this paper collects some datasets that is used frequently, and ways of error measurement are given. Finally, overall discussion is made including unsolved problems and currently challenging tasks.