Unable to reproduce reported metrics (mAP 54.16, HOTA 72.96, FSLA 54.17) with Swin-Laege backbone

[FightSH]

I'm attempting to reproduce the results reported in the paper using SWIN-LARGE as the backbone, but consistently fail to achieve the claimed performance metrics:
mAP: 54.16
HOTA: 72.96
FSLA: 54.17
Could you please clarify whether there are any special training configurations or tricks required for SWIN-LARGE to reach these numbers?

[Hellrogue]

The same situation，even worse than the origin baseline(2025 cvpr avis).