GeoDistill-Refine: Silhouette-First Distillation for Spacecraft Segmentation

A novel two-stage approach named GeoDistill-Refine has been developed to enhance spacecraft segmentation without the need for manual labels. This technique utilizes foundation segmentation models, particularly SAM 3, to create pseudo-masks that train a compact segmentation network. The outputs from the teacher model are stabilized by combining six fixed textual prompts through an unweighted 50% voting system. Initially, the student network focuses on learning the foreground silhouette, followed by refinement using signed-distance-field, skeleton, and area objectives based on the pseudo-mask. A sample-level gate, which considers prompt agreement, valid-prompt ratio, and pseudo-mask area plausibility, mitigates the influence of unreliable pseudo-geometry. GeoDistill-Refine shows improvements of 0.0456 in Image IoU and 0.1380 in Boundary F1 on the SpaceSense-Bench HJM lockbox set compared to a basic pseudo-label student. The research can be found on arXiv with the identifier 2608.07405.