A corner preserving surface inference algorithm using 3D convolution

Abstract A corner preserving 3D convolution operator is proposed. This vector convolution operator is a different version of Guy and Medioni's Diabolo field which is used for inferring surface normals from sparse and noisy 3D data. The new operator, named Quadratic field , is made by revolving a set of quadratic functions around the vertical axis. Other than the convolution operator, Guy and Medioni's method in surface normal inference is followed. Analysis and experimental results showed that the new operator performs better in recovering cornered surfaces from sparse 3D data.