Robustness of Input Features from Noisy Silhouettes in Human Pose Estimation

Silhouettes are frequently extracted and described to compose inputs for learning methods in solving human pose estimation problem. Although silhouettes extracted from background subtraction methods are usually noisy, the effect of noisy inputs to pose estimation accuracies is seldom studied. In this paper, we explore this problem. First, We compare performances of several image features widely used for human pose estimation and explore their performances against each other and select one with best performance. Second, iterative closest point algorithm is introduced for a new quantitative measurement of noisy inputs. The proposed measurement is able to automatically discard noise, like camouflage from the background or shadows. With the proposed measurement, we split inputs into different noise levels and assess their pose estimation accuracies. Furthermore, we explore performances of silhouette samples of different noise levels and compare with the selected feature on a public dataset: Human Eva dataset.