Surface Reconstruction of Point Clouds Based on Edge Occupancy Segmentation Distance Field

Implicit representation is an important method for 3D shape representation and reconstruction. The core of existing methods is to divide spatial points into internal points and external points to construct implicit fields. This strategy limits its ability to reconstruct open surfaces and multi-layer surfaces. In addition, existing methods rely on marching cubes to extract surfaces from implicit fields, which will lead to interpolation errors in nonlinear regions. To address the above two problems, an edge occupancy segmentation distance implicit field (EOSDF) method is proposed, which consists of two parts: an edge occupancy module and a segmentation distance module. The edge occupancy module adds segmentation points to two adjacent spatial points, and three points form a spatial edge. The occupancy relationship between the spatial edge and the target surface is directly analyzed, so that open surfaces and multi-layer surfaces can be reconstructed. The segmentation distance module calculates the distance between the spatial edge separated by the target surface according to the neighbor weights of the segmentation points, which is more conducive to determining the interpolation points and improving the surface reconstruction accuracy. Experimental results on the ShapeNetCore dataset show that the proposed method improves the chamfer distance and two F-score evaluation indicators by 13.4%, 7.8% and 1.7% respectively compared with GIFS. It can not only restore closed surfaces, but also effectively reconstruct open surfaces and multi-layer surfaces.