A Method of 3D Mesh Reconstruction Using a Single Color Image

Abstract

The invention belongs to the technical field of three-dimensional vision, and specifically relates to a method for reconstructing a three-dimensional mesh model of an object by using a single color image. The method of the present invention includes, for images, designing a multi-layer full convolution feature network for extracting the features of different levels of pictures; The ellipsoid is deformed to approximate the real shape; at the same time, the projection layer is designed to connect the image side and the 3D grid side, and an end-to-end neural network is trained under this framework, that is, given a color map, the corresponding 3D grid is output Model. This method has the advantages of smooth, complete and detailed 3D modeling results, and the reconstruction accuracy has also been effectively improved, which is very suitable for practical applications in industries such as virtual reality, animation games, and manufacturing.

Description

technical field [0001] The invention belongs to the technical field of three-dimensional vision, and in particular relates to a method for reconstructing a three-dimensional mesh model of an object by using a single color image. Background technique [0002] There are many representations of 3D data, including stereo raster, point cloud, mesh, etc. Stereo grid is a regular data structure, that is, objects are represented as N*N*N matrices, limited by resolution and expressive ability, this representation method usually lacks details and is difficult to represent complex shapes; 3D point cloud is a Irregular data structure, because there is no local connection relationship between points, the point cloud can only represent the approximate shape of the object, lacking the surface information of the object; the 3D grid is also an irregular data structure, composed of points, edges and Due to its light weight and rich shape details, it is more and more widely used in practical ...

AI Technical Summary

Problems solved by technology

However, this method is limited by the poor ability to expand category information, and the reconstructed model is severely lacking in details

[0007] The paper [3] reconstructs the three-dimensional raster. Due to the limitation of GPU memory and calculation amount, the output raster usually has a very low resolution.

Recently, the paper [4] proposed an octree representation method, which can reconstruct a model with a slightly higher resolution under the limited GPU memory budget, but the 3D grid is still not a commonly used representation method in the film and game industries.

[0008] The article [5] proposes a method to generate a point cloud from a single image. Since there is no local connection between the points, the degree of freedom of the position movement of the generated point is too high, so the reconstructed point cloud model lacks shape and surface details.

Images