Multiple-category, Multiple-instance Segmentation from RGB-D Data for Industrial Bin-picking

Bin-picking is a significant task in industrial scenes, with a challenging problem of segmentation. However, industrial scenes usually involve textureless objects with clutter and occlusion, further complicating recognition and segmentation. In this article, a new method is proposed to segment multiple-category, multiple-instance textureless objects with clutter and occlusion from RGB-D data for industrial bin-picking. The pipeline consists of a 2-D detector and a point cloud network. YOLOv3 is used as the 2-D detector to generate 2-D proposal and local point cloud. The modified RandLA-Net is proposed as the point cloud network to segment the local point cloud. A dataset of industrial pipes is labeled and used to train and test the proposed method. Ablation experiments show the effectiveness of the designs. Bin-picking experiments validate the capability of the method to apply in real-world systems.