MBBOS-GCN: minimum bounding box over-segmentation—graph convolution 3D point cloud deep learning model

Point cloud data with high accuracy and density is an important source for the depiction of real ground objects, there a broad research prospect using point directly 3D object detection recognition deep learning methods. However, many models in previous ignored structure information sampling randomness. To overcome this limitation, we proposed innovative model, namely, minimum bounding box over-segmentation–graph convolution network model (MBBOS-GCN) enhancing structural perception capability reduce In MBBOS-GCN, number points sampled used as scale, modified graph to collect from different scales. The divided into several small regions by algorithm, farthest (FPS) algorithm sample within each region experiments on classification semantic scene segmentation show that: (1) MBBOS-GCN has accuracy, which up 91.87% 89.5% ModelNet40 dataset ScanNet dataset, respectively; (2) provided good stability robustness little change under altering input data, slight loss value; (3) can be adapted complex scenes when reaches 97.53%. These superior performance provide effective support construction digital twin city background calibration multimode satellite feature inversion validation.