KITWARE INC — Department of Defense SBIR Phase I: OSD221-003

Description

Object detection datasets for overhead imagery are typically generated using bootstrapping methods to reduce annotator effort and cost. These methods iteratively train a detector from a limited set of user-provided and model-predicted labels. Such approaches bias detectors toward the initial object set, limiting their capacity to handle object variations or discover novel objects classes. MONET overcomes these limitations by inverting the problem, first learning to separate backgrounds from objects, then object classes from each other. Self-supervised learning clusters features of similar image patches, allowing for cluster-level, rather than patch-level labeling, significantly improving labeling efficiency. A theoretically grounded Extreme Value Machine (EVM), determines cluster similarities to already labeled clusters, enabling users to dynamically grow a class’s set of labeled clusters until the class is exhaustively labeled. In contrast, active learning only labels a subset of a given class for training. Further, characterizing detector embeddings with an EVM, MONET can provide theoretical estimates for the impact of individual false negatives on dataset labeling completeness. We evaluate using the xView dataset, measuring labeling completeness versus annotator time of MONET compared to typical bootstrapping, and the impact of bootstrapped false negatives on detector performance.