Evaluating Deep Contextual Description of Superpixels for Detection in Aerial Images

In several applications, the use of pixel contextual information has led to significant effectiveness results in remote sensing image classification tasks. In this paper, we investigate the use of deep superpixel context representation for detection in aerial images. Experimental results considering the widely used ISPRS Postdam and Munich Vehicle datasets demonstrate that max pooling and standard deviation are the most promising pooling methods, while the concatenation of contextual features from different pooling approaches leads to an effective characterization of superpixel contextual information in remote sensing images.