Bootstrapping CNNs for building segmentation in aerial imagery with depth

Detection of buildings and other objects from aerial images has various applications in urban planning and map making. For human experts, manual extraction of objects from aerial imagery is a time-consuming and expensive task. However, aerial images are often prone to varying lighting conditions, shadows and occlusions.Aerial images also differ quite significantly from region to region. In order to offer both robust and reliable building detection, we use Convolutional Neural Networks (CNN) to account for these variations. We train detection from RGB-D images to obtain a segmented mask by employing the CNN architecture DenseNet. To improve the performance of the model, we apply the statistical re-sampling technique called Bootstrapping. Through Bootstrapping, we demonstrate that more informative examples are retained that improves the results. Finally, the proposed methodology outperforms the non-bootstrapped DenseNet by utilizing only one-sixth of the original training data and it obtains a precision-recall break-even point at 95.10% score on our aerial imagery dataset.