Two-Stage Relation Constraint for Semantic Segmentation of Point Clouds

Abstract

Key to point cloud semantic segmentation is to learn discriminative representations involving of capturing effective relations among points. Many works add hard constraints on points through predefined convolution kernels. Motivated by label propagation algorithm, we develop Dynamic Adjustable Group Propagation (DAGP) with a dynamic adjustable scale module approximating distance parameter. Based on DAGP, we develop a novel Two Stage Propagation framework (TSP) to add intra-group and intergroup relation constraints on representations to enhance the discrimination of features from different group levels. We adopt well-appreciated backbone to extract features for input point cloud and then divide them into groups. DAGP is utilized to propagate information within each group in first stage. To promote information dissemination between groups more efficiently, a selection strategy is introduced to select group-pairs for second stage which propagating labels among selected group-pairs by DAGP. By training with this new learning architecture, the backbone network is enforced to mine relational context information within and between groups without introducing any extra computation burden during inference. Extensive experimental results show that TSP significantly improves the performance of existing popular architectures (PointNet, PointNet++, DGCNN) on large scene segmentation benchmarks (S3DIS, ScanNet) and part segmentation dataset ShapeNet.