In order to robustly and accurately restore the 3D vision information of an object from its two perspective views

by means of the new idea of randomly sampling the minimal redundant subset

by utilizing the data regularization technique

we develop a new algorithm

which can robustly and accurately recovers the 3D vision information of an object from its two perspective view data--the set of their feature point pairs. Random sampling can significantly reduce the sampling number of subset and make the good subset surely sampled. The redundant information contained in the minimal redundant subset can be efficiently used to check the validity and goodness of the sampled subset. The data regularization technique can greatly alleviate the numerical unstability generated from the ill posed property of the data. So

the algorithm is able to work well with high accuracy under very hard condition of heavy noise and high outlier rate. The experiments have demonstrated that the processed results are satisfactory.