Drone AI Trained on 34 Beluga Images Finds Calves Supervised Models Miss

Researchers from Florida Atlantic University (FAU) and Fisheries and Oceans Canada report that a semi-supervised approach, SEMI-DETR, can reduce expert annotation labor for beluga whale aerial surveys by five to ten times while outperforming supervised baselines. The study, published Sept. 4, 2026 in Frontiers in Marine Science, targets a key bottleneck: supervised detectors require extensive hand-labeled bounding boxes, and for belugas the labeling must be done by certified marine mammal biologists rather than crowdsourced volunteers. SEMI-DETR uses a teacher-student framework that generates pseudo-labels from unlabeled images and trains a student network under strong augmentation. The paper highlights three DETR-specific innovations, including stage-wise hybrid matching and cross-view query consistency, with especially large gains for calves.