Urban3A: Active Annotation Assisted Semantic Segmentation of Large-scale Urban Scenes

Semantic segmentation of point clouds is critically dependent on large volumes of precisely annotated data, a significant barrier especially in extensive urban environments. Prior research has primarily focused on object-level or indoor scenes, contingent upon densely annotated datasets for satisfactory results, rendering these methods less feasible for large urban applications. Addressing this, our study introduces a novel active learning-based selection framework, termed Urban3A, which leverages semi-supervised learning to mitigate the challenges of labeling in urban landscapes. Urban3A integrates a unique selection metric that combines basic geometric attributes, uncertainty, and deep learning features, preserving essential contextual information and enhancing the semantic depth of the data. This metric facilitates the identification of informative blocks and crucial points, markedly boosting the segmentation model's performance. Our annotation approach allows users to selectively annotate the most informative points, optimizing the training process. We validate our methodology using the demanding UrbanBIS dataset and benchmark against leading semi-supervised and active learning techniques. Results indicate that Urban3A achieves competitive performance with less than 0.1% of the labeled data, representing a significant advancement in efficient and cost-effective semantic segmentation of large-scale urban point clouds.