Weakly Supervised 3D Segmentation via Receptive-Driven Pseudo Label Consistency and Structural Consistency

As manual point-wise label is time and labor-intensive for fully supervised large-scale point cloud semantic segmentation, weakly method increasingly active. However, existing methods fail to generate high-quality pseudo labels effectively, leading unsatisfactory results. In this paper, we propose a segmentation framework via receptive-driven consistency structural mine potential knowledge. Specifically, three contrains: among different scales, structure between intra-class features class-level relation pair-wise categories. Three constraints are jointly used effectively prepares utilizes simultaneously stable training. Finally, extensive experimental results on challenging datasets demonstrate that our significantly outperforms state-of-the-art even achieves comparable performance the methods.