Weakly-Supervised Point Cloud Semantic Segmentation Based on Dilated Region

The escalating costs of labeling 3D point clouds have prompted researchers to investigate weakly supervised semantic segmentation. Current methods predominantly focus on designing network training processes, while somewhat effective in enhancing performance, overlook the inherent richness of point cloud information. This overlook often results in poor neural network generalization and a complex training process. We propose a novel approach, DR-Net, a weakly supervised semantic segmentation network based on Dilated Regional Feature Aggregation (DRFA), to address these issues. DR-Net directly employs the information of the point cloud itself, combines the advantages of local area feature aggregation at different scales, and uses feature richness for local enhancement to strengthen feature representation. Our U-shape-like overall framework further enhances network performance by fusing features at different levels after semantic querying. It is worth noting that the training process of DR-Net is similar to that of the fully supervised method, and commendable performance can be obtained without additional weak supervision tricks at the 1% labeled point, significantly reducing the training cost. DR-Net achieves competitive results on multiple benchmark datasets compared with state-of-the-art networks and demonstrates robust generalization performance in weakly labeled small sample scenarios. We also conducted a series of ablation experiments to verify DR-Net’s performance. Code is available at https://github.com/LujZhang/DR-Net.