A multi-task segmentation architecture with a Dilated ResNet encoder and dual decoders that jointly predict footprint masks and edge maps. Achieved SOTA 94% F1 on SpaceNet2 and WHU benchmarks.
Single-decoder segmentation networks trade off between crisp boundaries and solid masks — get one right and the other degrades. Building extraction needs both: solid masks for area, sharp edges for parcelization.
The "tuning fork" uses two decoders that share an encoder but predict complementary targets (mask + edge). A consistency loss forces them to agree on where the boundary sits, co-regularizing shape and edge.
Beat the prior SOTA (HRNet + BCE) by +1.6 F1 on SpaceNet2 and +2.1 on WHU, with stronger generalization across sensors.