Figure 2: Building map improvement following two rounds of training.

The network learns to group dense buildings when trained with small amounts of additional data that are similarly labeled. Extracted buildings are marked in transparent red.