SemanticFlow: Semantic Segmentation of Sequential LiDAR Point Clouds From Sparse Frame Annotations

Sequential point clouds acquired by light detection and ranging (LiDAR) technology provide accurate spatial information for environmental sensing. However, semantic segmentation of point cloud sequences relies on many manual point-wise annotations, which are error-prone and expensive. Existing mainstream weakly supervised methods tackle this by reducing the percentage of labeled points, but they are mostly designed for static indoor scenes and are hard to apply practically. From the viewpoint of realistic annotation procedures and the nature of point cloud sequences, this article proposes a novel semantic segmentation method, SemanticFlow, for LiDAR point cloud sequences using sparse frames with annotations. The proposed method achieves competitive performance compared with fully supervised methods. Specifically, we designed a bidirectional cross-frame pseudolabel propagation module that uses scene flow to learn the correlation and propagate pseudolabels across neighboring frames. In addition, a label refinement mechanism is proposed to select reliable pseudolabels for learning. Extensive experiments on SemanticKITTI, SemanticPOSS, and Synthia 4-D datasets demonstrate that our sparse frame annotation method is compatible with some fully supervised counterparts.