Active learning for annotation-free spacecraft detection via VLM knowledge distillation

Abstract

Reliable spacecraft detection remains difficult due to scarce labels in extreme visual conditions, while most existing annotation-free pipelines produce indiscriminate pseudo-labeling. We propose an annotation-free detection framework that combines Vision Language Model (VLM)-based Grounded SAM 2 pseudo-labeling with a hybrid active learning (AL) strategy, enabling a compact YOLOv8 detector to be trained without manually annotated labels. In each AL round, the framework selects informative images for pseudo-labeling based on model uncertainty, visual diversity, and random exploration. Experiments on the SPARK dataset show that one AL round achieves the best same-domain detection performance and outperforms both equal-volume random sampling and a larger YOLOv8s backbone. On the SPEED+ dataset, the proposed AL-guided model also shows better generalization ability, supporting the conclusion that selecting informative images is more effective than indiscriminately expanding pseudo-labels in annotation-free spacecraft detection. Additional comparisons with a CBAM-enhanced variant further suggest that lightweight attention offers only limited in-domain benefit while weakening cross-domain robustness.