In Object Centric OVD, the authors use the checkpoint trained from a subset of LMDet which has no overlap with any of the evaluation datasets. However, it seems that the checkpoint MDef_DETR_r101_epoch20.pth mentioned in Issue#22 is trained with the full dataset as it is the same checkpoint in evaluation. Thus the checkpoint used to generate class-agnostic proposals is needed to prevent leaking information. Please let me know if I missed anything. Thank you.