Conditional Generative Adversarial Networks for Data Augmentation and Adaptation in Remotely Sensed Imagery

摘要

The difficulty in obtaining labeled data relevant to a given task is among the most common and well-knownpractical obstacles to applying deep learning techniques to new or even slightly modified domains. The datavolumes required by the current generation of supervised learning algorithms typically far exceed what a humanneeds to learn and complete a given task. We investigate ways to expand a given labeled corpus of remotesensed imagery into a larger corpus using Generative Adversarial Networks (GANs). We then measure howthese additional synthetic data affect supervised machine learning performance on an object detection task.Our data driven strategy is to train GANs to (1) generate synthetic segmentation masks and (2) generateplausible synthetic remote sensing imagery corresponding to these segmentation masks. Run sequentially, theseGANs allow the generation of synthetic remote sensing imagery complete with segmentation labels. We applythis strategy to the data set from ISPRS' 2D Semantic Labeling Contest - Potsdam, with a follow on vehicledetection task. We find that in scenarios with limited training data, augmenting the available data with suchsynthetically generated data can improve detector performance.