owlv2 zeroshot performance on odinw13 dataset is inconsistency with paper record

[Yindong-Zhang]

we test owlv2-large-patch14-ensemble model on odinw13 dataset, only get 45.5 mAP in average, the same happens for finetuned and self-teaching model. quite lower than ~56 mAP in paper, any tricks or just release problem?

[Yindong-Zhang]

i use code from huggingface and get owlv2-base-patch16-ensemble around 46.5 on odinw13, which seems normal but owlv2-large performance is really pool.