Reinforcement Learning with Intelligent Context-based Exploration (RL-ICE)

This effort will extend the RL-ICE capability for low-shot object detection to operate with zero real data within the MWIR band (zero-shot learning). Our goal is to 1.) mature our software, 2.) maximize the performance benefit of synthetic data in training an object detector and, 3.) assess the trade-space of real vs synthetic data by comparing performance to a detector trained on fully real data. We expect the primary use case will be detection of a specific object of interest in motion imagery. We will collaborate with the Government and other performers to assess the relative performance of synthetic data and real data training methodologies.