Enhancing Weakly Supervised Point Cloud Semantic Segmentation via Dynamic Label Propagation and Pseudo-Label Optimization

• Published in: 2024 International Symposium on Intelligent Robotics and Systems (ISoIRS) pp. 138 - 142
• Main Authors: Jiang, Zefeng, Xiang, Pengcheng, Peng, Chengbin
• Format: Conference Proceeding
• Language: English
• Published: IEEE 14.06.2024
• Subjects: feature learning; Image processing; point cloud; Point cloud compression; pseudo-label; Representation learning; segmentation; Semantic segmentation; Smoothing methods; Supervised learning; Three-dimensional displays; weakly supervised
• DOI: 10.1109/ISoIRS63136.2024.00034

Point cloud semantic segmentation is a core task in the field of 3D computer vision, which aims to assign each point in a 3D point cloud to a specific category. Compared to semantic segmentation of 2D images, processing point clouds is more complex due to their often irregular structure and the large amount of spatial information they contain. However, at the same time, acquiring a large amount of accurately annotated 3D point cloud data is extremely time-consuming and costly. In practical applications, we may have a large amount of unannotated 3D point cloud data, but only a small portion is annotated. To overcome this challenge, we propose a weakly supervised learning approach for point cloud semantic segmentation, aimed at improving the performance of segmentation models by leveraging unannotated data. Initially, we partition the point cloud to obtain superpoints, and based on these superpoints, we construct a superpoint graph. Then, on the superpoint graph, we label a certain proportion of superpoints, and propagate pseudo-labels from these to surrounding superpoints with high confidence. Subsequently, we optimize the labels of superpoints with pseudo-labels to enhance the quality of the pseudo-labels. We also propose a multi-scale feature fusion method and a superpoint feature smoothing constraint to aid in the learning of superpoint features. Experiments demonstrate that our method achieves commendable performance on the publicly available datasets S3DIS and ScanNet.