Active Semi-Supervised 3d Semantic Segmentation for Sparsely Labeled Point Cloud

Abstract

Point cloud semantic segmentation is an imperative yet challenging task in various scenes. However, many deep-learning based methods heavily rely on large-scale point-wise labels, which are labor-intensive and time-consuming. To alleviate the need of dense annotation, recent methods are proposed to explore the semi-supervised 3D segmentation with limited randomly sampled annotations. However, random sampling usually leads to semantic repetition, thus is hard to improve performance. Therefore, the selection and utilization of labeled points remain unexplored and challenging task for semi-supervised segmentation. Different from existing one-time random sampling, this paper proposes a two-stage active semi-supervised segmentation method, exploring the learning-while-labeling framework. The semi-supervised learning stage aims at exploring the accurate classification boundary with limited annotation, and the active learning stage attempts to selected the most valuable points for segmentation. Concretely, in semi-supervised learning stage, a confidence-weighted dual-teacher network is proposed to learn a reliable feature space to supervise the student network, and the student network is forced to explore the accurate classification boundary under the combinational input perturbation. In the active learning stage, a hybrid uncertainty driven point-based active learning method is developed to determine the valuable unlabeled points for annotation. The unlabeled points are mapped into the feature space and measured by the hybrid uncertainty. The candidate points with high uncertainty are further processed by redundant point suppression module to remove redundant candidate points. The experiments on a real-world engine point cloud dataset and S3DIS dataset demonstrate that our method outperforms existing methods with fewer annotations.

Keywords: Semantic segmentation, semi-supervised learning, active learning, point cloud.