We augment the tensor voting framework with a datadriven multiscale scheme for reconstructing a multiresolution mesh from a noisy 3D point set. The augmentations are effective, automatic but very simple, consisting of surface saliency inference, scale segmentation, and data normalization. These data analysis steps enable tensor voting to operate at a single scale in each normalized data segment, by decoupling scale and smoothness control. They also guide tensor voting to reconstruct at optimal resolutions subject to the sampling theory. The output is a multiresolution mesh that captures large and small scale features faithfully, without using the maximum resolution everywhere in the domain. The augmented methodology is very robust in the presence of noisy and irregular samples, and non-trivial holes that cover large areas involving multiple-scale features.