Label-efficient geospatial point cloud semantic segmentation

Abstract

Recent advancements in point cloud semantic segmentation have consistently sur­passed previous state-of-the-art approaches. Nonetheless, the effectiveness of these models is heavily contingent upon the availability of extensive labeled data. The process of annotating large-scale geospatial point clouds, particularly those encom­passing multiple classes in urban environments, is exceptionally time-consuming and labor-intensive. This reliance on vast annotated datasets to achieve leading performance significantly hinders the practical applicability of large-scale point cloud semantic segmentation. Consequently, attaining promising results while sub­stantially minimizing labeling efforts is a crucial objective.