Building Detection and Segmentation Using a CNN with Automatically Generated Training Data

Abstract

Significantly outperforming traditional machine \nlearning methods, deep convolutional neural networks have gained increasing popularity in the application of image classification and segmentation. \nNevertheless, deep learning-based methods usually \nrequire a large amount of training data, which is \nquite labor-intensive and time-demanding. To deal \nwith the problem in generating training data, we \npropose in this paper a novel approach to generate image annotations by transferring labels from \naerial images to UAV images and refine the annotations using a densely connected CRF model with an \nembedded naive Bayes classifier. The generated annotations not only present correct semantic labels, \nbut also preserve accurate class boundaries. To validate the utility of these automatic annotations, we \ndeploy them as training data for pixel-wise image \nsegmentation and compare the results with the segmentation using manual annotations. Experiment \nresults demonstrate that the automatic annotations \ncan achieve comparable segmentation accuracy as \nthe manual annotations.