Satellite image providers usually present Rational Polynomial Coefficients (RPCs) as a user friendly solution for georeferencing of images. As RPCs are determined independently for each image scene, there are both absolute and relative georeferencing biases will in stereo scenes. Relative orientation of a stereo scene needs some conjugate image points. Speeded up robust features (SURF) operator is a powerful computer vision algorithm for image feature extraction and matching. Usually some of the features are not actually matched and are outliers. In this paper RANSAC+RPC algorithm is employed to simultaneously detect these outliers and modify the relative bias of RPCs. Our experiments on GeoEye-1 over Qom city and IRS-P5 over Rudehen district, both in central Iran, demonstrated the capability of our proposed algorithm. Though the RPC modification was done robustly for relative orientation of stereo scenes, yet improvement in the reconstructed 3D coordinates are in the range of sub-pixel accuracy. Our experiments demonstrate that the relative RPC shift and drift error will not cause any accuracy improvement in 3D reconstruction problem.