Cross-Dataset Pre-Training for Weakly Supervised Point Cloud Semantic Segmentation

With the rising of industries such as autonomous driving and robotics, as well as the maturity of various 3D sensing technologies, the application of processing point cloud data to perceive scenes has become increasingly common. This trend has made 3D point cloud semantic segmentation a hot topic in current research. However, despite the rich semantic information provided by point cloud data, fully annotating such large-scale and unstructured datasets remains highly challenging. To address this issue, weakly supervised learning methods have gained increasing attention, aiming to train networks with fewer labeled data. To further improve the performance of weakly supervised semantic segmentation of point clouds, a method that uses cross-dataset pre-trained models to assist weakly supervised learning is proposed. The Point Transformer architecture is employed for efficient multi-scale feature extraction, while the design of the dual-branch network facilitates the transfer of knowledge from different datasets, improving the performance of the weakly supervised learning process. Compared to the SQN weakly supervised learning method, this approach, with only 0.1% of the data labeled, increases the mIoU on the S3DIS dataset from 77.7% to 78.2% and on the Toronto3D dataset from 61.4% to 62.5%, significantly improving the performance of weakly supervised point cloud semantic segmentation.