Point cloud up-sampling method based on deep learning

Abstract

The invention discloses a point cloud up-sampling method based on deep learning. The point cloud up-sampling method comprises the steps of obtaining training data composed of a first number of sparseinput points and a second number of dense input points; constructing a deep network model for carrying out copying and curvature-based sampling operation on initial feature vectors extracted from a first number of sparse input points, obtaining a second number of intermediate feature vectors, performing splicing operation on each intermediate feature vector, inputting the intermediate feature vectors after the splicing operation into the multilayer perceptron, and determining a sampling prediction point based on the sampling feature vectors output by the multilayer perceptron; training a deepnetwork model until the target function determined by the sampling prediction point and the dense input point converges; and testing the deep network model to obtain point cloud data after sampling onthe test object. According to the method, the sparse point cloud can be converted into the dense point cloud based on curvature self-adaptive distribution, the object contour is accurately represented, and expression, rendering and visualization of three-dimensional data are better facilitated.

Description

[0001] technology neighborhood [0002] The invention relates to the field of computer vision technology, in particular to a point cloud upsampling method based on deep learning. Background technique [0003] With the popularity of depth cameras and lidar sensors, point clouds, as a simple and efficient representation of 3D data, have gradually attracted widespread attention from researchers. In recent years, researchers have used end-to-end neural networks to directly process raw point cloud data, and have made qualitative breakthroughs in visual tasks based on point cloud representations (eg, 3D object recognition and detection, 3D scene segmentation, etc.). However, the original point cloud data is usually generated by consumer-level scanning equipment, which has problems such as sparsity, incompleteness, and noise interference, which brings a huge challenge to point cloud semantic analysis. Therefore, in order to be used more efficiently for rendering, analysis or other p...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the technical problem of being unable to accurately characterize the contour shape of objects existing in related technologies, the present invention provides a point cloud upsampling method based on deep learning

Images