In this paper, we propose a dynamic fusion method that combines varifocal loss (VFL) and seesaw loss (SSL) to address the class imbalance problem in long-tail datasets for object detection models. The static combination of two loss functions makes the training unstable due to the noisy gradients of the SSL, which interrupt the IoU-aware classification flow and limit overall performance. To address this problem, we aim to effectively mitigate the biased learning problem in long-tailed datasets by maintaining the stability of IoU-aware classification during the early learning stages and gradually reflecting the calibration effect of the SSL in the latter stages of training. Finally, we validate unbiased detecting performance of the object detection on the LVIS dataset. To this end, we applied the proposed loss fusion method to the RT-DETRv2 model, resulting in 35.4% of bias mitigation for rare classes.
Kim et al. (Tue,) studied this question.