Previous research has shown the benefits of group equivariant convolutions for image recognition tasks. With this work we apply group equivariance to the segmentation of photovoltaic (PV) panel installations in aerial photography to determine whether the benefits translate to aerial photography segmentation. We create a custom annotation of PV panel installations in two Dutch cities using open access aerial photography. We show that group equivariant versions of traditional and residual

more »

... ional neural networks indeed perform at least as well as the traditional versions and provide better generalization. This work makes the following novel contribution; we show that group equivariant convolutions improve performance and generalization of fully convolutional networks applied to aerial photography segmentation. The remainder of this work is structured as follows. In Section 3 we first reiterate dilated and group equivariant convolutions and follow with a description