MODELING NON-OBJECTS AND NOVELTY FOR EFFICIENT TRAINING (MONET)

Overhead imagery object detection datasets are typically generated using bootstrapping methods that iteratively train a detector from an initial set of user-provided labels. Such approaches typically bias detectors toward the initial object set, limiting their capacity to handle object variations or discover novel and rare objects. In contrast, MONET learns a generalized object detector that can separate background from objects and discriminate object classes. Self-supervised learning computes an embedding space for clustering features of similar image patches, using extreme value theory (EVT)-based clustering. This allows for cluster level rather than patch level labeling, significantly improving labeling efficiency. EVT is also used to determine proposed object similarities to previously labeled objects, distinguishing between known and novel objects, and enabling users to exhaustively label all class instances with minimal user involvement. MONET uses EVEN, a new EVT-based method designed to address long-tail open set recognition and enable rare and novel object discovery. Finally MONET provides a theoretical estimate for the number of false negatives in the dataset. We present an integrated MONET system that can readily be applied to generate new datasets and demonstrate its effectiveness on NTM imagery.