Graph-SENet: An Unsupervised Learning-Based Graph Neural Network for Skeleton Extraction from Point Cloud
Extracting 3D skeletons from point clouds is a challenging task in computer vision. Most existing deep learning methods rely heavily on supervised data requiring extensive manual annotation. Consequently, re-labeling is often necessary for cross-category applications, while the process of 3D point cloud annotation is inherently time-consuming and expensive. Simultaneously, existing unsupervised methods often suffer from significant skeleton point deviations due to limited capabilities in modeling local structures. To address these limitations, we propose Graph-SENet, an unsupervised learning-based graph neural network method for skeleton extraction. This method integrates dynamic graph convolution with a multi-level feature fusion mechanism to more comprehensively capture local geometric relationships. Through a multi-dimensional unsupervised feature loss, it learns the structural representation of skeleton points, significantly improving the precision and stability of skeleton point localization under annotation-free conditions. Furthermore, we propose a graph autoencoder structure optimized by cosine similarity to predict topological connections between skeleton points, thereby recovering semantically consistent and structurally complete 3D skeleton representations in an end-to-end manner. Experimental results on multiple datasets, including ShapeNet, ITOP, and Soybean-MVS, demonstrate that Graph-SENet outperforms existing mainstream unsupervised methods in terms of Chamfer Distance and F1-score. It exhibits superior accuracy, robustness, and cross-category generalization capabilities, effectively reducing manual annotation costs while enhancing the completeness and semantic consistency of skeleton recovery. These results validate the application potential and practical value of Graph-SENet in 3D structure understanding and downstream 3D analysis tasks.