Military camouflaged object detection remains a challenging task due to the deliberate concealment of targets and the scarcity of annotated training data. In this paper, we propose an object-level distribution-aware synthetic data refinement framework that selects 3D-generated camouflage samples based on feature-space alignment with real and style-transferred images, measured by Mahalanobis distance in the detection feature space. The refined synthetic data is combined with real-image augmentation to form a series of training configurations evaluated across six lightweight YOLO variants targeting edge deployment on resource-constrained platforms. Experiments show that feature-based selection consistently outperforms using the full synthetic set, confirming that alignment quality matters more than quantity. Real-image style transfer drives the strongest per-class gains for visually distinctive but underrepresented classes, while selected 3D data contributes complementary geometric diversity. The combined strategy achieves the highest overall detection performance, with augmented configurations also converging faster and generalizing better to real-image validation data. Trade-off analysis across models reveals practical recommendations for balancing accuracy and computational cost under edge deployment constraints. Severely underrepresented categories remain a persistent bottleneck, highlighting the need for targeted data strategies beyond uniform augmentation.

Military camouflaged object detection;YOLO;data augmentation;synthetic data;style transfer;benchmark;lightweight;edge deployment;distribution-aware data refinement