Dual Adaptive Transformations for Weakly Supervised Point Cloud Segmentation

Weakly supervised point cloud segmentation, i.e. semantically segmenting a with only few labeled points in the whole 3D scene, is highly desirable due to heavy burden of collecting abundant dense annotations for model training. However, existing methods remain challenging accurately segment clouds since limited annotated data may lead insufficient guidance label propagation unlabeled data. Considering smoothness-based have achieved promising progress, this paper, we advocate applying consistency constraint under various perturbations effectively regularize points. Specifically, propose novel DAT (Dual Adaptive Transformations) weakly where dual adaptive transformations are performed via an adversarial strategy at both point-level and region-level, aiming enforcing local structural smoothness constraints on clouds. We evaluate our proposed two popular backbones large-scale S3DIS ScanNet-V2 datasets. Extensive experiments demonstrate that can leverage achieve significant performance gains datasets, setting new state-of-the-art segmentation.