System and method of open-world semi-supervised satellite object detection

Abstract

A system and method of open-world semi-supervised satellite object detection involves a machine learning engine configured with a training component and an inference component. The method of object detection detects objects in satellite imagery represents a solution for difficult challenges of arbitrary orientations, wide variation in object sizes, large number of densely packed objects, and highly complex background. A transformer network detects unknown objects in the satellite image. The transformer network includes a rotation-aware pyramidal pseudo-labeling operation that captures scale-specific pyramidal features at oriented box regions for pseudo-labeling unknown objects in the satellite image. A semi-supervised learning pipeline learns a new set of object classes to be detected. A prediction head outputs the satellite image annotated with a predicted object class for an unknown object. The inference component obtains object queries for a test satellite image, predicts labels for objects from known classes, predicts oriented boxes for the objects.