An Accurate Instance Segmentation Method for Point Clouds of Metallic Objects with Remarkable Pose Variations in Robotic Bin-Picking Scenes
Yilin Lu, Tingting Wang · 2024
This paper introduces a novel approach to instance segmentation tailored for industrial scenarios, focusing on the application to stackable aluminum alloy corner connectors in point cloud data. The proposed DEGCN (graph convolution network with directional encoding) employs directional encoding and region-growing techniques to address challenges posed by the unique characteristics of the industrial context. Extensive experiments conducted on a custom dataset and performance comparisons on the IPA dataset demonstrate the efficacy of DEGCN in achieving stable and accurate instance segmentation.