IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (Jan 2024)
Weakly-Supervised Semantic Segmentation of ALS Point Clouds Based on Auxiliary Line and Plane Point Prediction
Abstract
The segmentation of airborne lidar scanning (ALS) point clouds is one of the basic tasks in remote sensing field. The existed learning-based methods have acquired satisfactory performance, however, the training process requires a large amount of labeled data, which limits their wide application to some extent. Weakly-supervised learning of point clouds could achieve competitive results but far from perfect compared to fully supervised methods. In this article, we propose a weakly-supervised method for semantic segmentation of ALS point clouds based on learning of line and plane points. In addition, as basic feature units widely existing in large-scale scenes, both line and plane features could provide certain supervisory signals for training. Specifically, we extract line and plane points through two hand-crafted methods, which achieve offline line and plane point extraction. The extracted line and plane points are considered as additional supervisory signals, and consequently two auxiliary tasks could be constructed based on prediction of line and plane points, respectively. Furthermore, a multistage learning framework is proposed to embed the two auxiliary tasks into existing network, which makes the feature learning of network in the line-plane-object way. Moreover, we introduce an improved pseudolabel generation strategy based on prediction-consistency of contiguous models, which could generate high-quality pseudolabels so that to improve the performance of network. Extensive experiments conducted on three ALS point cloud datasets, i.e., ISPRS, LASDU, and DFC2019, demonstrate that the proposed method achieves considerable gains upon the baseline with respect to F1 score and overall accuracy under 0.1$\%$, 0.5$\%$, and 0.05$\%$ weak labels.
Keywords