Abstract
|
Preparation for stereoscopic 3D image shooting is tedious and time-consuming. Unsuccessful preparational setup prior to actual stereoscopic shooting can result in some ¡°bad¡± disparities such as vertical, scale, rotational differences in left and right images. The ideal case would be where we have zero disparities except the horizontal one, which is expensive and time-consuming to realize. In this paper, we investigated the margins of allowable stereoscopic geometry errors. In our experiments, subjects were asked to adjust the three kinds of disparities (vertical, scale, rotation), looking at images with ¡°bad¡± disparities, until they think it is comfortable to watch. We made Gaussian distribution over the experimental data and performed parameter estimation, including Maximum Likelihood Estimation and EM algorithm. From experimental result, rotational disparity showed the narrowest ¡°allowable¡± margin, while scale and vertical disparities showed some variabilities.
|
Keywords
|
Geometry error, Maximum likelihood estimator, Gaussian mixture model, EM algorithm, Stereoscopic errors
|