Dense Pyramid Network for Semantic Segmentation of High Resolution Aerial Imagery.

In this work, a dense pyramid network is proposed to provide fine classification maps of high resolution aerial images. The network applied densely connected convolutions to take full advantage of features and deepen the network without concerning the disappearance of gradients. Pyramid pooling module is introduced to bring flexible context information to the segmentation task and accomplish the fusion of multi-resolution features. Additionally, in order to preserve more information of multi-sensor data, group convolutions and channel shuffle operation are applied at the beginning of the network. We evaluate the dense pyramid network on the ISPRS Vaihingen 2D semantic labeling dataset, and the results demonstrate that the proposed framework exhibits better performance compared to the state of the art methods.