Hierarchical Intra-modal Correlation Learning for Label-free 3D Semantic Segmentation