Single-image three-dimensional point cloud model reconstruction method

Abstract

The invention relates to a single-image three-dimensional point cloud model reconstruction method, and the method comprises the steps: firstly carrying out the pre-training of a point cloud auto-encoder network, and obtaining a point cloud feature vector; constructing an image coding network to learn input image features, coding the input image features into an image feature vector with a fixed length, and then decoding the vector into a point cloud expression form by utilizing a decoder network; by minimizing the difference between the two feature vectors, the feature distribution of the input image is close to the feature distribution of the point cloud, so that a similar real point cloud model can be reconstructed, and the problem of uncertainty in three-dimensional reconstruction is solved. Moreover, the initial reconstruction point cloud is projected to a two-dimensional plane from different visual angles by utilizing a micro-projection module, a binary projection drawing is obtained, and the initial reconstruction point cloud is optimized by minimizing the difference between the initial point cloud projection drawing and a real projection drawing, so that a point cloud modelwith a detail structure is obtained.

Description

technical field [0001] The invention relates to the fields of computer vision and computer graphics, in particular to a three-dimensional reconstruction method based on a single two-dimensional image. Background technique [0002] In the real world, compared with massive and easy-to-obtain media data such as text and images, 3D models can better show the stereoscopic vision and detailed texture of objects, and are a data expression that is more in line with the human visual system. In addition, any With the development of hardware scanning equipment and professional modeling software, the abundance of 3D data, and the rapid development of computer storage space and computing power, 3D models are widely used in industrial design, urban planning, film and television education and other fields. However, hardware scanning equipment and professional modeling software such as 3DS Max, MAYA and other modeling software are expensive, and the threshold for use is high, which is not f...

AI Technical Summary

Problems solved by technology

[0005] The technical problem mainly solved by the present invention: Aiming at the current method based on image reconstruction mainly relying on 3D point cloud as the supervision method, a 3D point cloud model reconstruction method based on prior knowledge is provided, and the point cloud is obtained by pre-training the 3D point cloud model Prior knowledge not only reduces the generation of uncertain point clouds in the reconstruction process, but also combines multiple two-dimensional projection image supervision methods to make the reconstructed point cloud have the same appearance and contour details as the input. The invention reduces the need for real three-dimensional The dependence of model supervision improves the accuracy of 3D model reconstruction and makes 3D model reconstruction more extensive

Images