Supervoxel-based and Cost-Effective Active Learning for Point Cloud Semantic Segmentation

Recent successes in point cloud semantic segmentation heavily rely on a large amount of annotated data to train a deep neural network. Furthermore, three dimensional (3D) point cloud data generally has no order and sparsity, and a point cloud often includes more than ten thousand points, thus increasing difficulties of point cloud annotation. To reduce the huge annotation efforts, we propose a supervoxel-based and cost-effective active learning pipeline which aims to select only uncertain and diverse segmented regions for annotation. To better exploit annotating budget, we first change the annotating units from a point cloud scan to segmented regions through two unsupervised methods. We further propose to leverage point cloud intensity when calculating the segmented region information for encouraging region diversity. Extensive experiments show that our approach greatly outperforms previous active learning methods, and we achieve up to 90% performance of a fully supervised trained deep neural network by only using 3% labeled data compared to 100% on SemanticKITTI dataset.