FeaturePointNet: A Semi-Supervised Geometric-Aware Graph Transformer for Dominant Point Extraction from Building Outlines

Extracting dominant points from complex building outlines is an important task in cartography and building footprint simplification. This study proposes a semi-supervised, geometric-aware graph-Transformer framework for automatic dominant point selection from vector point sequences. The framework combines geometric rules, topological constraints, and data-driven learning to facilitate dominant point detection. A graph network is used to capture local neighborhood structures, while a Transformer models global morphology and long-range dependencies. To improve detection stability, Gaussian scale space is introduced to analyze contours across multiple smoothing scales, and multi-scale geometric features are extracted to distinguish structural changes from local noise. Experimental results demonstrate that the proposed method achieves the best overall performance in building footprint simplification. Indicates that the proposed method provides more accurate, robust, and geometrically stable dominant points for subsequent building reconstruction.