Robust multi-view triangulation algorithms usually rely on an empirical reprojection error threshold to identify and remove the outliers. The selection of such threshold is critical to both the quantity of successfully reconstructed scene point and its accuracy. Based on the analysis of the noise in feature point localization and the geometry of epipolar transfer, the uncertainty propagation model in epipolar transfer was derived. A novel noise scale estimation approach based on kernel density estimation was proposed and the estimated noise scale was further incorporated into robust state-of-the-art multi-view triangulation algorithm. Experimental results demonstrate that the proposed method is able to obtain accurate estimation of noise scale and to improve the 3D reconstruction quality of multi-view triangulation algorithm significantly.