Instance Segmentation of Low-texture Industrial Parts Based on Deep Learning

Abstract

The instance segmentation of low-texture industrial parts is important for robot grasping operations in scattered environments. However, most of the current deep learning methods for instance segmentation rely heavily on the RGB information of the scene, which limits their application in low-texture scenes and some scenes where RGB information cannot be obtained; there are fewer point cloud datasets for industrial parts. The deep learning method based on point cloud is not ideal for the segmentation of scattered and stacked industrial parts with complex shapes. In this paper, a dataset for industrial parts is generated in a physical simulation environment, and a deep learning method for instance segmentation of low-texture industrial parts based on the point cloud is proposed. The simulation dataset experiment verifies that the method can achieve instance segmentation of low-texture industrial parts in scattered stacking scenes, and has strong robustness to point clouds with inconsistent densities and noise.