To address the National Geospatial-Intelligence Agency (NGA) need for overhead imagery analysis algorithms that provide uncertaintymeasures for object recognition and aggregation, Intellisense Systems, Inc. (ISS) proposes to develop a new Variational Object Recognition andGrouping Network (VORGNet) system. It is based on the innovation of implementing a Bayesian convolutional neural network (CNN) thatdetects objects in overhead images with uncertainty measures, followed by object grouping and group classification, taking into accountvarious uncertainties. The object detection uncertainty is estimated by variational inference, conveniently integrated with the dropoutregularization technique widely used in deep learning. The detected objects are grouped by fuzzy clustering. Each group is characterized by itsmembers and the spatial configuration and classified by supervised machine learning and domain knowledge. Monte Carlo (MC) simulationsincorporate the object detection uncertainty and the group membership uncertainty to quantify the overall uncertainty in aggregation. InPhase I, ISS will prove VORGNets feasibility by demonstrating its ability to obtain uncertainty measures for an object detection CNN andperform uncertainty-aware object aggregation on publicly available overhead imagery datasets. In Phase II, ISS will refine VORGNetsperformance on various overhead imagery datasets and develop test metrics jointly with the government for performance evaluation.
