Convolutional network three-dimensional model reconstruction method based on multi-view cost volume

Abstract

The invention discloses a convolutional network three-dimensional model reconstruction method based on multi-view cost volume, and the method comprises the steps: extracting a feature map of input multi-view picture data through a coding network sharing the weight; according to a parallel plane of a reference camera viewpoint, the feature maps are distorted to different depths, and a feature volume corresponding to each feature map is obtained; fusing the plurality of feature volumes into one cost volume by using variance-based cost measurement; and carrying out noise reduction processing on the obtained cost volume by adopting 3D grid reasoning, reconstructing a final voxel, and completing convolutional network three-dimensional model reconstruction based on the multi-view cost volume. According to the method, three-dimensional grid reasoning is used, geometric constraints are added, and high-quality output is generated; each view angle is used as a primary reference view angle, so that the independence between the final result and the input is ensured, and the rationality of the network is improved; a new depth value is selected such that the shape of the resulting result coincides with the shape used to supervise.

Description

technical field [0001] The invention relates to a three-dimensional model reconstruction method, in particular to a convolutional network three-dimensional model reconstruction method based on multi-view cost volume. Background technique [0002] In computer vision, 3D reconstruction refers to the process of depth data acquisition, preprocessing, point cloud registration and fusion, surface generation, etc., to describe the real scene into a mathematical model that conforms to computer logic. In recent years, with the explosive growth and availability of 3D data, and the emergence of deep learning methods, 3D reconstruction has become a hot field. The focus of 3D reconstruction technology is how to obtain the depth information of the target scene or object. Under the condition that the depth information of the scene is known, the three-dimensional reconstruction of the scene can be realized only through the registration and fusion of the point cloud data. The deep applicat...

AI Technical Summary

Problems solved by technology

However, the differentiable arg-min operation is easily affected by multiple modes, and the implementation effect is not ideal

Images