Enhancing Object Detection Algorithm for Size-Insensitive Performance

Detecting tiny objects remains a challenge in computer vision, particularly in aerial imagery, due to limited resolution and scale variation. Conventional Intersection-over-Union (IoU)-based loss functions often lead to scale-sensitive optimization, degrading detection consistency across object sizes. Normalized Wasserstein Distance for Small Object Detection (NW-SOD) has addressed scale sensitivity, but still faces limitations in capturing the full distributional characteristics for robust multi-scale detection. This study proposes Bhattacharyya-distance-based Multi-scale Object Detection (B-MOD), a loss function that achieves balanced detection performance across varying object scales without modifying the network architecture. B-MOD utilizes the Bhattacharyya distance to measure distributional similarity between predicted and ground truth bounding boxes, enabling more accurate regression and confidence estimation. Comparative evaluations conducted on the DOTA dataset show that B-MOD demonstrates better performance than existing distribution-aware methods across diverse object scales. Results show improvements over both IoU-based metrics and NW-SOD across different object scales, with gains observed in small object detection accuracy. These findings indicate that distributional similarity-based loss functions can improve multi-scale object detection performance.