Low-Shot Detection in Remote Sensing Imagery

The National Geospatial-Intelligence Agency (NGA) ingests and analyzes raw imagery from multiple sources to form actionable intelligenceproducts that can be disseminated across the intelligence community (IC). To effectively meet these demands NGA must continue to improveits automated and semi-automated methods for target detection and classification. Of particular concern is furthering NGA's ability to identifyand locate rare intelligence targets of interest within large search regions in remote sensing imagery. This is known as the low-shot problem.Deep learning methods continue to demonstrate state-of-the-art performance in the problem domains of computer vision and machinelearning, including for low-shot detection. However, only recently have these deep networks been applied to the remote sensing domain.Challenges exist for adapting such deep learning approaches to the remote sensing domain including lack of large corpuses of training dataand significant differences in image appearance when compared to traditional deep learning problems and datasets. Toyon proposes theresearch and development a low-shot detection algorithm that relies on deep learning to; produce a feature representation specific to theremote sensing domain, leverage 3D class and image information jointly, generate realistic hallucinated imagery of low-shot classes.