Label optimization point cloud instance segmentation method

Abstract

The invention belongs to the technical field of point cloud instance segmentation, and discloses a label optimization point cloud instance segmentation method. The method comprises the steps: firstly carrying out the feature extraction of a point cloud through a graph convolution neural network; establishing an instance label matrix for the training set, and performing label propagation on the instance label matrix by using a label propagation algorithm to obtain an optimized instance label matrix; and finally, performing instance segmentation on the point cloud by combining the label matrix and the optimized instance label matrix. According to the method, the similarity relationship of the points in the point cloud is fused, so that the supervision of the label matrix has better distinguishing representation, and a better segmentation result is obtained. According to the method, the global shape information and the local feature information of the large-scene point cloud model are considered, and the effect of integrating the global information is achieved; a label propagation algorithm is used for optimizing a label matrix, and the point cloud segmentation result is improved through the combined action of original data set labels and optimized labels.

Description

technical field [0001] The invention belongs to the technical field of point cloud instance segmentation, and in particular relates to a tag-optimized point cloud instance segmentation method. Background technique [0002] At present, the instance segmentation of point cloud refers to identifying the similarities and differences between different objects under a certain threshold range through the collection and integration of model features and the use of corresponding algorithms. Instance segmentation mainly includes two main tasks: feature extraction of point cloud and parameter adjustment of loss function for instance segmentation. However, in the existing point cloud instance segmentation methods, there are mainly three directions: the first is to voxelize the point cloud, and use the regularity of voxels in the three-dimensional space to perform convolution, but the disadvantage lies in the calculations consumed by voxelization. The volume is large, and the accuracy o...

AI Technical Summary

Problems solved by technology

[0004] The difficulty of solving the above problems and defects is: the scale of the data set is an unavoidable problem for deep learning, the small amount of data leads to overfitting, and the model cannot be applied to other data

The existing deep learning framework for segmenting point clouds restricts the upper limit of segmentation accuracy, and can only improve performance through various other optimization methods, and the improvement is limited

Images