Technical validation of label-efficient blueberry instance segmentation under extreme annotation scarcity

Pseudo-label selection becomes particularly fragile under extreme annotation scarcity because confidence-based filtering can reject low-confidence masks that remain valid instance hypotheses. We address this problem with a label-efficient blueberry instance-segmentation pipeline that combines image-adaptive teacher-confidence filtering with a geometry-based Rescue mechanism. The adaptive branch accommodates image-level variation in teacher scores, while Rescue selectively reconsiders rejected candidates using task-specific mask-shape cues based on circularity and solidity. Under a 1% labeled protocol on the Mega Blueberry benchmark, Adaptive+Rescue achieves a mean Mask mAP50 of 0.777 ± 0.021 across three labeled-subset realizations and three student-training seeds. Relative to the matched adaptive-confidence branch, Rescue yields a mean paired improvement of +0.0241 Mask mAP50, with a crossed-bootstrap 95% confidence interval of [0.0020, 0.0416]. A ground-truth-linked audit further shows that 95.08% of the 2,478 Rescue-only masks match annotated instances at IoU \(\ge 0.50\) . These results show that task-specific mask geometry can provide a complementary reliability signal for pseudo-label selection when confidence filtering alone is insufficient, supporting label-efficient blueberry segmentation under extreme annotation scarcity.