Solving the confusion of body sides problem in two-dimensional human pose estimation
I would also like to recognize SAFRAN-France (MORPHO) for sponsoring a fundamental part of this work under the Research Award Program.
Abstract
In this thesis, we address the problem of two-dimensional human pose estimation (HPE) from a single viewpoint. While many approaches to estimate the 2D human pose from a single viewpoint exist, the estimated joints' locations with respect to the viewpoint are often disregarded. This limits the overall accuracy of localizing the human body parts. To address this limitation, we define a novel problem in 2D HPE: the Confusion of Body Sides (CBS). We show the CBS problem in many 2D HPE approaches as well as in the state-of-the-art methods. In order to overcome the CBS problem, we introduce SHAPE: Smart Human Articulated Pose Estimation. We demonstrate how SHAPE can be plugged into a 2D HPE algorithm to solve the CBS problem. We report our qualitative and quantitative results on our proposed challenging dataset: "Humans AUC" as well as on two popular HPE benchmark datasets: "KTH Multiview Football dataset II" [1] and "Image Parsing" [2]. Our approach is shown to make a notable 2D HPE approach [3] viewpoint-invariant and enhance the accuracy by 20% on average.