Recognition of a person from gait has been a recent focus in computer vision. It is one biometric source that can be acquired at a distance. At this nascent stage of gait recognition research, the pertinent research questions are those related to understanding the limits of gait recognition and the quantitative study of the various factors effecting gait. However, performances of contemporary algorithms have been confounded by errors in the extracted silhouettes, which has been the low-level representation of choice. In this work, (i) we present to the research community a segmentation "ground truth" research resource consisting of a set of manually specified part-level silhouettes for 70 subjects from the recently formulated Gait Challenge database, under different conditions involving change in surface, shoe-type, and time; a total of about 8000 manual silhouettes. (ii) We expound an HMM Eigen Stance modelbased silhouette reconstruction method to correct for common errors in silhouette detection arising from shadows and background subtraction. And (iii) using these "cleaned" silhouettes and the manual silhouettes we show that the effects of various factors such as surface, time, and shoe on gait recognition are not due to poor silhouette quality. In fact, the recognition performance actually drops with the use of "clean" silhouettes because of removal of correlation in the error pixel patterns.