A Study of the Impact of Attention Mechanisms on Feature Correlation Learning for Building Extraction Models

Buildings are the fundamental units of civilization, deeply ingrained into urban elements and synonymous with progress. Therefore, comprehending buildings through extracting their footprint data is vital for assessing and influencing societal advancement. Unlike the direct approach of recent research, we recognize that using identical or related representation is undoubtedly potential through feature correlation learning. In this study, we examined how plugging the current attention mechanisms in two fundamental building extraction models led to notable enhancements in their performance. Specifically, we achieved a 1.25% increase in the IoU score for CrossGeoNet and a 2.7% improvement for HiSup.