Enhanced Semantic Segmentation for Large-Scale and Imbalanced Point Clouds

Semantic segmentation of large-scale point clouds is of significant importance in environment perception and scene understanding. However, point clouds collected from real-world environments are usually imbalanced and small-sized objects are prone to be under-sampled or misclassified due to their low occurrence frequency, thereby reducing the overall accuracy of semantic segmentation. In this study, we propose the Multilateral Cascading Network (MCNet) for large-scale and sample-imbalanced point cloud scenes. To increase the frequency of small-sized objects, we introduce the semantic-weighted sampling module, which incorporates a probability parameter into the collected data group. To facilitate feature learning, we propose a Multilateral Cascading Attention Enhancement (MCAE) module to learn complex local features through multilateral cascading operations and attention mechanisms. To promote feature fusion, we propose a Point Cross Stage Partial (P-CSP) module to combine global and local features, optimizing the integration of valuable feature information across multiple scales. Finally, we introduce the neighborhood voting module to integrate results at the output layer. Our proposed method demonstrates either competitive or superior performance relative to state-of-the-art approaches across three widely recognized benchmark datasets: S3DIS, Toronto3D, and SensatUrban with mIoU scores of 74.0\%, 82.9\% and 64.5\%, respectively. Notably, our work yielded consistent optimal results on the under-sampled semantic categories, thereby demonstrating exceptional performance in the recognition of small-sized objects.