Next Century Corporation proposes the development of the Computer Vision Synthetic Image Generation Harness for Training and Testing (CV SIGHTT), a prototype system that learns to train a generation-based image recognition system for low-shot detection in the remote sensing domain. CV SIGHTT will use Deep Evolution Strategies to select the procedure for synthetic image generation, choosing the one that maximizes object detection performance. Evolution Strategies learning incorporates delayed reward, fully explores the parameter space, and represents the procedure in a human-legible form. In addition, we will research and apply a Deep Convolutional Neural Network (DCNN) that provides a value function to allow faster and more accurate exploration, building from insights from recent training methods such as Asynchronous Actor-Critic Agents (A3C) and other methods, to create Deep Evolution Strategies. The object detection task for CV SIGHTT will be recognizing uncommon targets in remote imagery, given 3D models and a set of domain imagery.
