The performance of deep learning-based object detection methods is heavily dependent on the characteristics of training data. However, real-world detection datasets often exhibit severe imbalance in object-scale distributions, resulting in insufficient supervision for object sizes with limited instances. To address this issue, we propose a scale-aware mosaic augmentation algorithm and a generalized scale-adaptive intersection over union (GSIoU)-based varifocal loss (VFL) function. The proposed scale-aware mosaic augmentation method alleviates object-scale imbalance by explicitly regulating scale distributions during training sample construction, overcoming the tendency of conventional mosaic augmentation to preserve inherent dataset imbalance. Furthermore, to handle the limitation of existing IoU-based quality targets that impose relatively large penalties for small objects, we replace the localization quality target in VFL with GSIoU, thereby enabling more consistent classification performance across object scales. We evaluate the effectiveness of the proposed method by applying the proposed method to RT-DETRv2 and conducting experiments on the HRSC2016-MS dataset. Experimental results demonstrate that the proposed method improves the overall average precision (AP) by 1.11AP over the baseline, with a particularly notable improvement of 16.43AP in small-object average precision, confirming that the proposed approach achieves balanced detection performance across object scales.
Kim et al. (2026) studied this question.