Low-Shot Alternate Viewpoint Analogies

Overhead imagery analysts spend their time scouring aerial and satellite imagery looking for objects based on example images. Collectingexample imagery of our adversaries newest military hardware is often challenging. The only examples we have may be from open sourceintelligence at air shows or military parades or from limited clandestine collections. Often, only a small set of ground-based imagery isavailable. Based on their experience and training, it is up to the imagery analyst to surmise what that object would look like from overheadwhile searching through imagery to find it. Low-Shot Alternate Viewpoint Analogies (LAVA) will automate this process. We will develop a deeprepresentation and viewpoint analogy generator network to detect objects in overhead images. LAVAs viewpoint analogy generator will learnthe overhead appearance of objects from ground-based imagery. Using state-of-the-art low-shot learning, we will extend the classifier to newobject classes using as few as one ground-based image. With LAVAs automatic object detection, the analysts work will be sped up, and theirtime will be freed for higher-level cognitive tasks. Our goal is for LAVA to have a positive impact in diverse areas including disaster recovery,customs enforcement, and nuclear nonproliferation.