Reconstructing the missing dimension: From 2D to 3D human pose estimation

We address the task of estimating a 3D human pose given a 2D pose estimate. A promising approach was presented by Taylor in 2000. The approach needs no learning and is based on a simple principle, namely exploiting the foreshortening information of projected limbs. Though it received only little attention due to two severe restrictions: it uses an unrealistic camera model - the scaled orthographic projection - and yields no unique solution. We show how to overcome both restrictions. We first present an extension of Taylor's original method to a realistic camera model, i.e. perspective projections. Since the method still does not yield an unique solution but a whole set of pose candidates we show how to reduce this set of candidates further by exploiting anatomical constraints and joint angle probabilities. The method is evaluated on the public available TUM kitchen dataset and shows that the average reconstructed joint angle error is in the range of 4.5°-8.2° even for camera views showing strong perspective effects.