Progressive indoor scene point cloud up-sampling method based on sparse semantic guidance

Abstract

The invention provides an indoor scene-oriented sparse semantic guidance progressive point cloud up-sampling method, and solves the problems of sparsity, high noise, irregular distribution and the like of point clouds collected by a 3D (three-dimensional) sensor. A progressive up-sampling mechanism is designed based on sparse tensor, and a progressive point generation module is constructed in the progressive up-sampling mechanism and used for iteratively generating and pruning intermediate point clouds. By introducing a plurality of generative deconvolution operations and adopting self-adaptive pruning strategies of different coefficients, the target point cloud can be reconstructed step by step, and meanwhile, a local fine-grained structure is effectively reserved. In addition, a sparse semantic embedding module is introduced into each layer of the encoder, so that the extraction capability of point features is enhanced. In order to further improve the feature expression effect, the invention further provides a multi-scale sparse semantic embedding module, and fusion features are refined in space and channel dimensions. A wide range of experiments on a plurality of indoor scene data sets show that the method is continuously superior to an existing most advanced method in most evaluation indexes, and excellent performance is shown.

Images