This aerial image dataset consists of more than 22,000 independent buildings extracted from aerial images with 0.0075 m spatial resolution and 450 km^2 covering in Christchurch, New Zealand. The most parts of aerial images are down-sampled to 0.3 m ground resolution and cropped into 8,189 non-overlapping tiles with 512* 512. These tiles make up the whole dataset. They are split into three parts: 4,736 tiles for training, 1,036 tiles for validation and 2,416 tiles for testing.
We employed training set and validation set to update network weight \omega and architecture parameters (\alpha, \gamma) in search phase, respectively. When the search is over, we retrained the emitted architecture on the whole training set from scratch, and finally evaluated the trained network on the testing set.