KDC-3D: KD-Tree-Driven Clustering for Fast Instance Segmentation in Industrial Point Clouds

Segmentation of instances in industrial point clouds is a critical task for applications such as robot bin-picking, autonomous navigation and human-robot collaboration. Specifically, segmentation of numerous similar objects that are occluded from each other is a major challenge, especially in robotic bin-picking tasks. In a typical bin-picking scenario where multiple identical objects are stacked together, we ignore semantic information and focus on instance segmentation. In this paper, we propose KDC-3D, a novel KD-tree driven clustering algorithm for fast and accurate segmentation of instances in industrial point clouds.KDC-3D utilizes Dynamic Graph Convolutional Networks (DGCNN) to extract point features and predict geometric centers, and then uses a KD-Tree based clustering algorithm to assign each remaining point to the nearest geometric center in the feature embedding space. Experimental results show that KDC-3D outperforms the most available methods in terms of segmentation accuracy and computational efficiency, especially in industrial segmentation scenarios. Our method greatly reduces the inference time while maintaining high precision and recall, making it suitable for real-time applications.