Abstract
This paper studies the problem of polygonal mapping of buildings by tacklingthe issue of mask reversibility that leads to a notable performance gap betweenthe predicted masks and polygons from the learning-based methods. We addressedsuch an issue by exploiting the hierarchical supervision (of bottom-levelvertices, mid-level line segments and the high-level regional masks) andproposed a novel interaction mechanism of feature embedding sourced fromdifferent levels of supervision signals to obtain reversible building masks forpolygonal mapping of buildings. As a result, we show that the learnedreversible building masks take all the merits of the advances of deepconvolutional neural networks for high-performing polygonal mapping ofbuildings. In the experiments, we evaluated our method on the two publicbenchmarks of AICrowd and Inria. On the AICrowd dataset, our proposed methodobtains unanimous improvements on the metrics of AP, APboundary and PoLiS. Forthe Inria dataset, our proposed method also obtains very competitive results onthe metrics of IoU and Accuracy. The models and source code are available athttps://github.com/SarahwXU.