11-30, 13:50–14:10 (Asia/Seoul), Seoul Archive
The proliferation of high-resolution satellite imagery has dramatically transformed urban planning, disaster response, and environmental monitoring domains. A persistent challenge, however, is the precise extraction and delineation of building footprints, as traditional methods often yield results with irregular shapes and misalignments. To address these limitations, we present CornerRegNet, a pioneering deep neural network designed to predict oriented building corners and subsequently regularize building footprints derived from satellite imagery.
Central to CornerRegNet's design is its unique ability to harness oriented building corners as primary anchoring points. This approach significantly streamlines and refines the shape of the extracted building footprints. Using advanced multi-modal autoencoders, CornerRegNet adeptly embeds a blend of corner features, their inherent orientations, and overarching semantic context. This intricate fusion of features allows the network to capture micro-level details and macro-level building structural context.
Enhancing the accuracy of corner prediction, the model incorporates active rotating filters. These filters, combined with uniquely devised equivariant convolutions, ensure that the predicted corners are not only spatially correct but also maintain their intended orientation concerning the building's overall layout.
A pivotal component of CornerRegNet is the integrated Graph Convolution Network (GCN). This GCN is instrumental in iteratively refining the positions of corners from the preliminary footprints. Further, the in-built shape regularizer within the GCN aids in achieving architecturally consistent and coherent footprints, effectively minimizing typical extraction anomalies.
In comparative evaluations, CornerRegNet stands out, consistently outperforming contemporary models. The network's resilience in managing and regularizing building shapes is a distinguishing attribute, even under challenging scenarios marked by occlusions.
In conclusion, CornerRegNet represents a unique advancement in building footprint extraction from satellite imagery. Its unique combination of corner orientation prediction and shape regularization techniques sets new standards for accuracy, efficiency, and architectural interpretability.
Dr. Gunho Sohn is a distinguished Associate Professor and Department Chair at the Department of Earth and Science and Engineering at York University. He is also the Inaugural Director of the Mobility Innovation Centre (MOVE). Dr. Sohn's innovative research in urban digital twinning, which combines computer vision, and machine learning, has garnered acclaim for making urban environments smarter, safer, and more sustainable.