Detecting building changes with off-nadir aerial images

The tilted viewing nature of the off-nadir aerial images brings severe challenges to building change detection (BCD) problem: mismatch nearby buildings and semantic ambiguity facades. To tackle these challenges, we present a multi-task guided network model, named as MTGCD-Net. proposed model approaches specific BCD problem by designing three auxiliary tasks, including: (1) pixel-wise classification task predict roofs facades buildings; (2) an for learning roof-to-footprint offsets each account misalignment between roof instances; (3) identical matching flow bi-temporal problem. These tasks provide indispensable complementary parsing information. predictions are finally fused main branch with multi-modal distillation module. train test models images, create new benchmark dataset, BANDON. Extensive experiments demonstrate that our achieves superior performance over previous state-of-the-art competitors.