Low-Shot Detection in Remote Sensing Imagery

Toyon Research Corporation proposes to research and develop algorithms for low-shot object detection, adapting popular techniques to address the complexities inherent in ATR for remote sensing. Traditional object detection algorithms rely on large corpora of data which may not be available for more exotic targets (such as foreign military assets), and therefore, traditional Convolutional Neural Network (CNN) based approaches must be adjusted for this low-shot detection problem. Toyons proposed effort to address the difficulties of low-shot detection in remote sensing consists of: (i) the development of a feature representation for remote sensing imagery, (ii) incorporation of additional data modalities (such as text description of targets) to improve detection, (iii) state-of-the-art methods of exemplar synthesis from existing images of other targets classes, (iv) Image Matching networks for determining the visual similarity of candidate detections with a small list of exemplars and (v) external Memory Augmentation of Neural Networks to extend the above algorithms to adapt to new, unseen target classes (zero-shot detection).