Example-based pose estimation in monocular images using compact fourier descriptors

Automatically estimating human poses from visual input is useful but challenging due to variations in image space and the high dimensionality of the pose space. In this paper, we assume that a human silhouette can be extracted from monocular visual input. We compare the recovery performance of Fourier descriptors with a number of coefficients between 8 and 128, and two different sampling methods. An examplebased approach is taken to recover upper body poses from the descriptors. We test the robustness of our approach by investigating how shape deformations due to changes in body dimensions, viewpoint and noise affect the recovery of the pose. The average error per joint is approximately 16-17° for equidistant sampling and slightly higher for extreme point sampling. Increasing the number of descriptors does not have any influence on the performance. Noise and small changes in viewpoint have only a very small effect on the recovery performance but we obtain higher error scores when recovering poses using silhouettes from a person with different body dimensions.