While many works focus on 3D reconstruction from images, in this paper, we\nfocus on 3D shape reconstruction and completion from a variety of 3D inputs,\nwhich are deficient in some respect: low and high resolution voxels, sparse and\ndense point clouds, complete or incomplete. Processing of such 3D inputs is an\nincreasingly important problem as they are the output of 3D scanners, which are\nbecoming more accessible, and are the intermediate output of 3D computer vision\nalgorithms. Recently, learned implicit functions have shown great promise as\nthey produce continuous reconstructions. However, we identified two limitations\nin reconstruction from 3D inputs: 1) details present in the input data are not\nretained, and 2) poor reconstruction of articulated humans. To solve this, we\npropose Implicit Feature Networks (IF-Nets), which deliver continuous outputs,\ncan handle multiple topologies, and complete shapes for missing or sparse input\ndata retaining the nice properties of recent learned implicit functions, but\ncritically they can also retain detail when it is present in the input data,\nand can reconstruct articulated humans. Our work differs from prior work in two\ncrucial aspects. First, instead of using a single vector to encode a 3D shape,\nwe extract a learnable 3-dimensional multi-scale tensor of deep features, which\nis aligned with the original Euclidean space embedding the shape. Second,\ninstead of classifying x-y-z point coordinates directly, we classify deep\nfeatures extracted from the tensor at a continuous query point. We show that\nthis forces our model to make decisions based on global and local shape\nstructure, as opposed to point coordinates, which are arbitrary under Euclidean\ntransformations. Experiments demonstrate that IF-Nets clearly outperform prior\nwork in 3D object reconstruction in ShapeNet, and obtain significantly more\naccurate 3D human reconstructions.\n