Few-Shot Aircraft Door Detection with Visibility-Aware Geometric Regularization

Aircraft-door detection is a key perception component for automated passenger boarding bridges, but its performance can degrade substantially under limited annotations, illumination changes, partial occlusion, and viewpoint variation. To address this problem, we propose a multi-constraint geometric regularization (MGR) framework for few-shot aircraft-door detection. MGR represents each door using four corner landmarks and a derived center landmark, and explicitly couples appearance learning with door-frame geometry. Specifically, projectively valid diagonal-incidence constraints are combined with visibility-aware and geometry-gated symmetry and orthogonality regularization, while multi-scale RoI feature aggregation and landmark-guided refinement provide complementary appearance cues. Experiments use ResNet-18, ResNet-34, and ResNet-50 with matched target-domain training-instance budgets and support sets, while distinguishing box-only from landmark supervision. With ResNet-50, MGR achieves AP₅₀ values of 77.90%, 82.10%, 85.00%, and 86.70% under 15-, 20-, 25-, and 30-shot settings, respectively, giving an average observed margin of 1.10 percentage points over SGFNet, the strongest competing method in the main comparison tables. Rankings are mixed with ResNet-18 and ResNet-34. Ablations indicate additional gains from the combined geometric objective, although refinement contributes a larger observed AP₅₀ increase in the ResNet-50 comparison and individual geometric terms are not uniformly beneficial across backbones. These results indicate that explicitly modeling reliable door-frame geometry can provide useful structural supervision for data-efficient aircraft-door detection, particularly under challenging visual conditions.