DAS-PAFPN: Collaborative Dual Attention Network for Instance Segmentation in Dense Building Scenarios

Abstract

Accurate instance segmentation of densely packed buildings in high-resolution remote-sensing imagery remains challenging because narrow inter-building gaps, weak boundary contrast, and cluttered urban backgrounds frequently cause adjacent buildings to merge into single masks. We propose the Dual-Attention Synergistic Path Aggregation Feature Pyramid Network (DAS-PAFPN), a structure-semantics collaborative feature-fusion neck for RTMDet-Ins. Rather than fusing multiple sensor modalities, DAS-PAFPN integrates two complementary representations within a single-image feature hierarchy. The Residual Column Spatial Self-Correction (RCSSC) module replaces conventional fusion units in the top-down and bottom-up pathways to encode anisotropic row-column structural priors and preserve rectilinear boundary cues. The Hybrid Pooling Attention (HPA) module then operates on the neck outputs to aggregate complementary pooled context and recalibrate feature responses before dynamic mask prediction. We further introduce Instance Count Deviation (ICD) as an instance-count-based complement to COCO AP metrics for diagnosing under- and over-segmentation in dense scenes. Experiments on the newly constructed PDBD and the public TCBD and UVBD benchmarks demonstrate consistent improvements in segmentation accuracy, boundary fidelity, and instance separability. Additional comparisons with recent state-of-the-art approaches on the challenging TCBD-Dense subset confirm the competitiveness of DAS-PAFPN under strict mask-overlap and recall-oriented criteria. Overall, the proposed framework provides an effective structure-semantics collaborative approach to densely packed building instance segmentation.

Keywords

Instance Segmentation / Dense Buildings / Remote Sensing Imagery / Attention Mechanism / RTMDet / Topological Correctness

Rights & permissions

Higher Education Press 2026

Just Accepted

This article has successfully passed peer review and final editorial review, and will soon enter typesetting, proofreading and other publishing processes. The currently displayed version is the accepted final manuscript. The officially published version will be updated with format, DOI and citation information upon launch. We recommend that you pay attention to subsequent journal notifications and preferentially cite the officially published version. Thank you for your support and cooperation.