Aaron Reite, National Geo-Spatial Intelligence Agency
gtc-dc 2019
The need for large volumes of data labeled for specific tasks is perhaps the most common obstacle for successfully applying deep learning. We’ll discuss how to apply deep learning on satellite imagery when labeled training data is limited, or even non-existent. First, we’ll show how a state-of-the-art unsupervised algorithm can learn a feature extractor on the noisy and imbalanced xView data set that readily adapts to several tasks: visual similarity search that performs well on both common and rare classes; identifying outliers within a labeled data set; and learning a class hierarchy automatically. Then we’ll demonstrate how GANs can be applied to jointly model images and their labels to generate an expanded training set, and how this may be applied to the data set from ISPRS’s Potsdam 2D Semantic Labeling Contest to improve a follow-on vehicle detection task when training data is limited.