Computer vision and machine-learning techniques have revolutionized our world in recent years. However, for robust operation, these techniques need very large amounts of data and annotation. This requirement places real limits on their utility in critical situations where access to data may be limited.