Volumetric Human Reconstruction from a Single Depth Map

Abstract

We present an efficient approach to reconstruct a human body from a single depth map, captured by a commercial depth camera or a stereo depth sensor. The underlying idea is to predict the rear side depth map through the deep network because the rear side depth map tends to symmetric to the front depth map and the shape variation is lesser than the front. One the rear side depth map is predicted, we construct a signed distance volume and extract a human as the form of 3D meshes through the Marching Cubes method. We experimentally show that the proposed method can effectively predict the rear side depth map.