Mask2Former–FGMBRM: A Feature-Guided Building Segmentation Framework for Automated 3D Urban Environment Generation from Satellite Imagery

Building-footprint errors can propagate into geospatial assets generated from remote-sensing imagery. This study evaluated Mask2Former–FGMBRM, which applies feature-guided residual refinement to Mask2Former query-mask predictions and integrates extracted footprints with documented terrain data and geospatial processing for footprint-based urban asset generation. Six models were evaluated on 518 WHU held-out images using common final references and metric definitions while retaining model-specific pipelines. The tune-selected Mask2Former baseline and FGMBRM achieved AP values of 72.03% and 73.49%, respectively, a 1.47-percentage-point improvement. FGMBRM also achieved 89.01% binary IoU and 91.86% object F1, leading these measures and AP among the six configurations, whereas SegFormer had the highest boundary F1. Controlled ablations showed metric-dependent trade-offs, with simpler variants leading selected outcomes and no consistent advantage from query-conditioned guidance. The original INRIA comparison remained limited. In a separate query-capacity diagnostic, AP at common maxDets = 100 increased from 20.66% to 23.96%, while retaining more outputs increased recall and false positives. The results support footprint-based urban asset generation and visualization, with remaining dense-scene and geometric limitations.