Abstract
Building vectorization from remote sensing imagery is a critical task challenged by three persistent issues: distinguishing buildings from complex surroundings, adapting to vast scale variations, and delineating intricate boundaries. While existing methods have addressed these issues with specific techniques, they often lack a unified framework that systematically integrates these capabilities. Inspired by the hierarchical process of human cartography, which progresses from a coarse overview to fine-grained refinement, we propose a guiding cognitive model for network design. This model decomposes the task into three sequential principles: Necessity, establishing foundational scene understanding with a sufficient receptive field; Diversity, adapting feature representation to targets of varying sizes; and Sensitivity, focusing on critical boundary details. Based on this model, we introduce TriadMapper, a network for high-precision building vectorization. TriadMapper implements this cognitive hierarchy through three integrated components: (i) it strategically adopts a large-kernel mechanism, which we term the Context-Aware Selective Kernel (CA-SK), to establish the necessary broad contextual awareness; (ii) it introduces a novel Global-Local Perception (GLP) module that adaptively fuses multi-scale features to address the diversity of building scales and shapes; and (iii) it proposes a novel Scene Adaptive Optimization (SAO) module that uses deformable attention for sensitive boundary localization. Furthermore, our Mask-guided Vertex Selection and Optimization (MVSO) strategy employs adaptive geometric thresholds for high-quality polygon generation. Comprehensive experiments on the AICrowd benchmark demonstrate state-of-the-art performance, achieving 81.6% AP and 0.680 PoLiS. TriadMapper validates that a framework guided by cognitive principles can be a robust and precise solution for building vectorization. Code is available at: https://github.com/ParkourX/TriadMapper.
Funding Information
  • National Natural Science Foundation of China (42201491, 42371461)

This publication has 61 references indexed in Scilit: