Self-supervised Transformer for 3D point clouds completion and morphology evaluation of granular particle

Abstract

Determining the morphology characteristics of particles using 3D point cloud is promising and crucial for the quality inspection of granular materials. However, it remains challenging due to the cumbersome process and incomplete 3D point clouds obtained from laser scanning of particles. In this study, a novel intelligent method, named self-supervised transformer-based encoder and decoder model for granular materials (SSPoinTr-GM), is developed for the automatic completion of partially occluded 3D point clouds and morphology characteristics evaluation. The complete cloud points of 100 cobble and 100 gravel particles are first scanned to establish a benchmark 3D point cloud dataset. To form partial point clouds for training, the complete point cloud is divided into global seed points by the farthest point sampling (FPS) method and the local cloud points around each seed point by the k-nearest neighbor method. Then, the seed points and their local cloud points are randomly removed to generate partial cloud points as input, training the encoder and decoder in a self-supervised way with the original complete point cloud as ground truth. Experiments are conducted to validate the effectiveness of the novel method compared with four existing completion baselines based on the 3D point cloud dataset. The results indicate that the CD1 loss of the completed particles by the proposed method is, on average, 49.05 % lower than that of existing baselines. Additionally, the error rate of the calculated morphology characteristics of the completed particles is, on average, 66.06 % lower than that of the partial point clouds.