3d Segmentation Methods of Archaeology Sites Using Dynamic Graph CNN and Transformer Architecture

In this paper 3d segmentation methods of irregular point clouds are presents to decipher the structure of archaeological sites of the Bronze Age in the Southern Trans-Urals. The first method for 3d semantic segmentation is based on a multimodal dynamic graph, which is created by recalculating the graph's Kirchhoff matrix in each special convolutional layer of the neural network. The second method for 3d instance segmentation is based on the architecture of a deep neural network in the form of improved Mask3D transformer. The introduced modifications into the transformer architecture made it possible to significantly improve the confidence of instance segmentation and obtain a dense, complete and homogeneous point cloud for processing from various depth sensors. For computer simulation, the most promising methods for 3d semantic segmentation and 3d instance segmentation were selected. In this paper computer simulation was carried out for various methods of 3d segmentation, the results obtained were presented and discussed, and the dependence of the accuracy segmentation from 3d up-sampling and augmentation procedure of point clouds was studied.