A weakly supervised framework for real-world point cloud classification

Real-world point cloud objects pose great challenges in point cloud classification as objects acquired by scanning devices from real-world scans are often cluttered with background, and are partial due to occlusions as well as reconstruction errors. In the literature, few works tackle the problem of real-world point cloud classification while existing methods require fully point-level annotated training samples. However, large-scale dense point-level foreground–background labeling for real-world point clouds is a labor-intensive and time-consuming job. Leveraging two auxiliary modules, called semi-supervised point-level pseudo labels generation and noise-robust multi-task loss, the framework can integrate well with existing supervised point cloud classification network. A relational graph convolutional network on the local and non-local graph (PointRGCN) is first proposed to predict point-level foreground–background pseudo labels for each object with sparse ground-truth point-level foreground–background labels in training datasets. Then, a weakly supervised classification network, which combines with an auxiliary foreground–background segmentation branch, is employed to classify real-world point clouds. To cope with noise-containing point-level foreground–background labels generated above, a noise-robust multi-task loss is proposed to train the network accurately. Experimental results show that the performance of the proposed framework which is trained with only 1% point-level labels is comparable with many popular or state-of-the-art fully supervised methods. The source code will be available at .