Research On High Relief Generation From 3D Model

As a popular art form,relief has been widely used in daily life and industry.Traditional modeling of reliefs relies on a large amount of human interactions and is timeconsuming.In recent years,many works researchers have been conducted on automatic relief modeling,significantly improving the quality and efficiency of relief creation.However,these works mainly focus on the modeling of bas-reliefs,and few efforts has been put forward on high relief generation.In this work,we study the problem of high relief modeling,where two different data types,e.g.3D mesh and 3D point cloud are given as references.(1)High-relief modeling from 3D mesh through differential deformation.We propose a novel approach for high-relief modeling,which benefits from Laplace-based mesh deformation.Given a 3D mesh as input,we first compress the object from a predefined viewing direction.Then we select a set of handle points on the back side of the object based on the computation of mean curvature and normal angle.After that,we optimize the depth field by solving a bi-Laplacian-based linear system.As a result,the deformed object is ensured to be attached to the background.In case the object is composed of topology-disconnected parts,we further define connection handles to link the parts together.The deformation is ensured to be transferred to each part of the object so that the optimized high-relief keeps the geometric information and visual appearance of the input.Experimental results show that the proposed method can deal with different kinds of input objects and output high reliefs with high quality.(2)High-relief modeling from 3D point cloud through neural network.To improve the modeling efficiency,we propose a neural based method for high-relief modeling,which automatically outputs depth coordinates of the input point cloud.First,we build high-quality high relief dataset to enable neural network training.To test the ability of neural modeling,we construct two types of end-to-end networks.By comparisons,we finally choose the MLP-based network.Experimental results show that the MLP-based network is efficient in generating high reliefs from point clouds with reasonable depth orders and rich geometric details.