Weakly-Supervised Point Cloud Segmentation Combining Pseudo Label Generation and Noise Label Learning

Aiming at the problem that current deep learning-based point cloud segmentation methods require a large amount of dense point-level labeled training data, a weakly-supervised point cloud segmentation framework based on pseudo label generation and noisy label learning is proposed. Firstly, based on the non-local similarity of point clouds, we employ local and non-local graphs with multiple types of edges to model the point cloud. Secondly, we introduce a semi-supervised relational graph convolutional network to generate high-quality pseudo labels for incompletely labeled point cloud training data. Finally, to tackle the noisy labels in generated pseudo labels, a progressive noise-robust loss function is proposed to accurately train the point cloud segmentation network on noisy pseudo labeled data. The proposed framework is evaluated on public point cloud segmentation datasets, ShapeNet Part and S3DIS. Experiments show that without adding extra computation in inference stage, the proposed method can achieve segmentation accuracy 0.1% lower than fully supervised method in ShapeNet Part dataset with 10% labels, and 5.2% lower than fully supervised method in S3DIS dataset with 1% labels.