Adaptive Grid Selection Training Strategy for Tiny Object Detection

Detecting small objects and managing occlusions remain persistent challenges in object detection tasks, particularly in complex scenarios with diverse environments or densely packed scenes. These challenges often result in poor detection accuracy and limited model generalization across varying datasets. To address these challenges, we propose two novel methods: Object-Oriented Cutout (OOC) and Selective Grid Loss Function (SGLoss). OOC enhances training data diversity while preserving small object integrity by strategically applying object-aware augmentation. SGLoss optimizes grid cell allocation dynamically based on object size and aspect ratio, ensuring the generation of high-quality positive samples while reducing inconsistencies in grid-level assignments, and further enabling precise alignment between anchors and ground truth boxes for improved detection accuracy across scales. Our proposed methods achieve a 0.6% improvement in mean Average Precision (mAP) compared to the YOLOv5 baseline, with significant gains observed across small and occluded object categories in datasets such as VisDrone 2019, DOTA, MS COCO 2017, PASCAL VOC, and SODA10M. These contributions advance the field of object detection by addressing critical limitations and provide a foundation for future research in diverse domains, including autonomous driving, aerial imagery analysis, and real-time surveillance systems.