Gaussian Dynamic Convolution for Semantic Segmentation in Remote Sensing Images

Different scales of the objects pose a great challenge for segmentation remote sensing images special scenes. This paper focuses on problem large-scale variations target via dynamical receptive field deep network. We construct Gaussian dynamic convolution network by introducing layer to enhance image understanding. Moreover, we propose new pyramid pooling (GPP) multi-scale object segmentation. The proposed can expand size and improve its efficiency in aggregating contextual information. Experiments verify that our method outperforms popular semantic methods large datasets, including iSAID LoveDA. conduct experiments demonstrate works more effectively than other convolutional layers.