Small-object detection is a challenging problem especially in aerial imagery due to large variations in object scale, high density, and limited spatial resolution. The Focus-and-Detect (F&D) framework addresses this problem by dividing the process into two stages: a Focus stage that predicts object-dense regions, and a Detect stage that performs fine-grained detection within these focal crops. In this study, we present HalF&D, a parameter-efficient variant of the F&D framework that integrates the HaLViT (Half of the Weights are Enough) mechanism into both stages. HaLViT reduces parameters by sharing a single projection matrix and its transpose across bottleneck blocks within each residual stage. Applied to ResNet-50 in the Focus stage and ResNeXt in the Detect stage, HalF&D achieves significant compression while maintaining usable accuracy on the VisDrone dataset. Specifically, the Focus network achieves 47.9% mAP@50 with a 47% parameter reduction, and the Detect network reaches 32.2% mAP with more than 70% fewer parameters. These results demonstrate that parameter sharing can substantially reduce model size while preserving the spatial precision and recognition ability in HalF&D.
