Zachary Phaneuf, Cloudstone Innovations; Nathan Toner, Cloudstone Innovations
https://doi.org/10.64861/UATR5819
Keywords: Detection, localization, machine learning, vision transformer, wide field of view, surveillance
Abstract:
The proliferation of large satellite constellations in low Earth orbit (LEO) is driving demand for automated, high-accuracy detection of the satellite streaks they leave in wide-field astronomical imagery—a core capability for space-domain awareness (SDA), and one where labeled data are scarce and the targets are sparse, faint, and extremely elongated. We study how to turn a frozen self-supervised vision transformer into an oriented streak detector for this setting. Holding the backbone (a DINO-style ViT-S/16), the training data, and the box-quality loss fixed, we vary only the detection head and compare two designs. A DETR-style query head with one-to-one bipartite matching is diagnosed, per image, to fail—it hits a threshold-insensitive recall ceiling, over-predicts on empty and dense frames, and reads out a collapsed, image-independent cross-attention, even though linear probes show the frozen features encode position cleanly. Replacing it with a dense, anchor-free oriented-box head—task-aligned label assignment, probabilistic-IoU regression, an orientation-aware angle loss, and a non-max-suppression-free method—removes the ceiling. Under a unified center-matching protocol on synthetic validation imagery, we find order-of-magnitude tighter center and orientation localization than the query-head baseline.
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications