A novel clustering approach demonstrates that similarity-based representation learning combined with density-based aggregation is a promising strategy for point cloud segmentation in highly granular particle detectors.