The fundamental matrix is a two-view tensor that plays a central role in Computer Vision geometry. We address its robust estimation given correspondences between image features. We use a non-parametric estimate of the distribution of image features, and then follow a probabilistic approach to select the best possible set of inliers among the given feature correspondences. The use of this perception-based \acontrario principle allows us to avoid the selection of a precision threshold as in RANSAC, since we provide a decision criterion that integrates all data and method parameters (total number of points, precision threshold, number of inliers given this threshold). Our proposal is analyzed in simulated and real data experiments; it yields a significant improvement of the ORSA method proposed in 2004, in terms of reprojection error and relative motion estimation, especially in situations of low inlier ratios.